1 /* 2 * Copyright (c) 2005, 2014, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "compiler/compileLog.hpp" 27 #include "libadt/vectset.hpp" 28 #include "opto/addnode.hpp" 29 #include "opto/callnode.hpp" 30 #include "opto/castnode.hpp" 31 #include "opto/cfgnode.hpp" 32 #include "opto/compile.hpp" 33 #include "opto/convertnode.hpp" 34 #include "opto/locknode.hpp" 35 #include "opto/loopnode.hpp" 36 #include "opto/macro.hpp" 37 #include "opto/memnode.hpp" 38 #include "opto/narrowptrnode.hpp" 39 #include "opto/node.hpp" 40 #include "opto/opaquenode.hpp" 41 #include "opto/phaseX.hpp" 42 #include "opto/rootnode.hpp" 43 #include "opto/runtime.hpp" 44 #include "opto/subnode.hpp" 45 #include "opto/type.hpp" 46 #include "runtime/sharedRuntime.hpp" 47 48 49 // 50 // Replace any references to "oldref" in inputs to "use" with "newref". 51 // Returns the number of replacements made. 52 // 53 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) { 54 int nreplacements = 0; 55 uint req = use->req(); 56 for (uint j = 0; j < use->len(); j++) { 57 Node *uin = use->in(j); 58 if (uin == oldref) { 59 if (j < req) 60 use->set_req(j, newref); 61 else 62 use->set_prec(j, newref); 63 nreplacements++; 64 } else if (j >= req && uin == NULL) { 65 break; 66 } 67 } 68 return nreplacements; 69 } 70 71 void PhaseMacroExpand::copy_call_debug_info(CallNode *oldcall, CallNode * newcall) { 72 // Copy debug information and adjust JVMState information 73 uint old_dbg_start = oldcall->tf()->domain()->cnt(); 74 uint new_dbg_start = newcall->tf()->domain()->cnt(); 75 int jvms_adj = new_dbg_start - old_dbg_start; 76 assert (new_dbg_start == newcall->req(), "argument count mismatch"); 77 78 // SafePointScalarObject node could be referenced several times in debug info. 79 // Use Dict to record cloned nodes. 80 Dict* sosn_map = new Dict(cmpkey,hashkey); 81 for (uint i = old_dbg_start; i < oldcall->req(); i++) { 82 Node* old_in = oldcall->in(i); 83 // Clone old SafePointScalarObjectNodes, adjusting their field contents. 84 if (old_in != NULL && old_in->is_SafePointScalarObject()) { 85 SafePointScalarObjectNode* old_sosn = old_in->as_SafePointScalarObject(); 86 uint old_unique = C->unique(); 87 Node* new_in = old_sosn->clone(sosn_map); 88 if (old_unique != C->unique()) { // New node? 89 new_in->set_req(0, C->root()); // reset control edge 90 new_in = transform_later(new_in); // Register new node. 91 } 92 old_in = new_in; 93 } 94 newcall->add_req(old_in); 95 } 96 97 // JVMS may be shared so clone it before we modify it 98 newcall->set_jvms(oldcall->jvms() != NULL ? oldcall->jvms()->clone_deep(C) : NULL); 99 for (JVMState *jvms = newcall->jvms(); jvms != NULL; jvms = jvms->caller()) { 100 jvms->set_map(newcall); 101 jvms->set_locoff(jvms->locoff()+jvms_adj); 102 jvms->set_stkoff(jvms->stkoff()+jvms_adj); 103 jvms->set_monoff(jvms->monoff()+jvms_adj); 104 jvms->set_scloff(jvms->scloff()+jvms_adj); 105 jvms->set_endoff(jvms->endoff()+jvms_adj); 106 } 107 } 108 109 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word, int mask, int bits, bool return_fast_path) { 110 Node* cmp; 111 if (mask != 0) { 112 Node* and_node = transform_later(new AndXNode(word, MakeConX(mask))); 113 cmp = transform_later(new CmpXNode(and_node, MakeConX(bits))); 114 } else { 115 cmp = word; 116 } 117 Node* bol = transform_later(new BoolNode(cmp, BoolTest::ne)); 118 IfNode* iff = new IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN ); 119 transform_later(iff); 120 121 // Fast path taken. 122 Node *fast_taken = transform_later(new IfFalseNode(iff)); 123 124 // Fast path not-taken, i.e. slow path 125 Node *slow_taken = transform_later(new IfTrueNode(iff)); 126 127 if (return_fast_path) { 128 region->init_req(edge, slow_taken); // Capture slow-control 129 return fast_taken; 130 } else { 131 region->init_req(edge, fast_taken); // Capture fast-control 132 return slow_taken; 133 } 134 } 135 136 //--------------------copy_predefined_input_for_runtime_call-------------------- 137 void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) { 138 // Set fixed predefined input arguments 139 call->init_req( TypeFunc::Control, ctrl ); 140 call->init_req( TypeFunc::I_O , oldcall->in( TypeFunc::I_O) ); 141 call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ????? 142 call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) ); 143 call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) ); 144 } 145 146 //------------------------------make_slow_call--------------------------------- 147 CallNode* PhaseMacroExpand::make_slow_call(CallNode *oldcall, const TypeFunc* slow_call_type, address slow_call, const char* leaf_name, Node* slow_path, Node* parm0, Node* parm1) { 148 149 // Slow-path call 150 CallNode *call = leaf_name 151 ? (CallNode*)new CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM ) 152 : (CallNode*)new CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), oldcall->jvms()->bci(), TypeRawPtr::BOTTOM ); 153 154 // Slow path call has no side-effects, uses few values 155 copy_predefined_input_for_runtime_call(slow_path, oldcall, call ); 156 if (parm0 != NULL) call->init_req(TypeFunc::Parms+0, parm0); 157 if (parm1 != NULL) call->init_req(TypeFunc::Parms+1, parm1); 158 copy_call_debug_info(oldcall, call); 159 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON. 160 _igvn.replace_node(oldcall, call); 161 transform_later(call); 162 163 return call; 164 } 165 166 void PhaseMacroExpand::extract_call_projections(CallNode *call) { 167 _fallthroughproj = NULL; 168 _fallthroughcatchproj = NULL; 169 _ioproj_fallthrough = NULL; 170 _ioproj_catchall = NULL; 171 _catchallcatchproj = NULL; 172 _memproj_fallthrough = NULL; 173 _memproj_catchall = NULL; 174 _resproj = NULL; 175 for (DUIterator_Fast imax, i = call->fast_outs(imax); i < imax; i++) { 176 ProjNode *pn = call->fast_out(i)->as_Proj(); 177 switch (pn->_con) { 178 case TypeFunc::Control: 179 { 180 // For Control (fallthrough) and I_O (catch_all_index) we have CatchProj -> Catch -> Proj 181 _fallthroughproj = pn; 182 DUIterator_Fast jmax, j = pn->fast_outs(jmax); 183 const Node *cn = pn->fast_out(j); 184 if (cn->is_Catch()) { 185 ProjNode *cpn = NULL; 186 for (DUIterator_Fast kmax, k = cn->fast_outs(kmax); k < kmax; k++) { 187 cpn = cn->fast_out(k)->as_Proj(); 188 assert(cpn->is_CatchProj(), "must be a CatchProjNode"); 189 if (cpn->_con == CatchProjNode::fall_through_index) 190 _fallthroughcatchproj = cpn; 191 else { 192 assert(cpn->_con == CatchProjNode::catch_all_index, "must be correct index."); 193 _catchallcatchproj = cpn; 194 } 195 } 196 } 197 break; 198 } 199 case TypeFunc::I_O: 200 if (pn->_is_io_use) 201 _ioproj_catchall = pn; 202 else 203 _ioproj_fallthrough = pn; 204 break; 205 case TypeFunc::Memory: 206 if (pn->_is_io_use) 207 _memproj_catchall = pn; 208 else 209 _memproj_fallthrough = pn; 210 break; 211 case TypeFunc::Parms: 212 _resproj = pn; 213 break; 214 default: 215 assert(false, "unexpected projection from allocation node."); 216 } 217 } 218 219 } 220 221 // Eliminate a card mark sequence. p2x is a ConvP2XNode 222 void PhaseMacroExpand::eliminate_card_mark(Node* p2x) { 223 assert(p2x->Opcode() == Op_CastP2X, "ConvP2XNode required"); 224 if (!UseG1GC) { 225 // vanilla/CMS post barrier 226 Node *shift = p2x->unique_out(); 227 Node *addp = shift->unique_out(); 228 for (DUIterator_Last jmin, j = addp->last_outs(jmin); j >= jmin; --j) { 229 Node *mem = addp->last_out(j); 230 if (UseCondCardMark && mem->is_Load()) { 231 assert(mem->Opcode() == Op_LoadB, "unexpected code shape"); 232 // The load is checking if the card has been written so 233 // replace it with zero to fold the test. 234 _igvn.replace_node(mem, intcon(0)); 235 continue; 236 } 237 assert(mem->is_Store(), "store required"); 238 _igvn.replace_node(mem, mem->in(MemNode::Memory)); 239 } 240 } else { 241 // G1 pre/post barriers 242 assert(p2x->outcnt() <= 2, "expects 1 or 2 users: Xor and URShift nodes"); 243 // It could be only one user, URShift node, in Object.clone() instrinsic 244 // but the new allocation is passed to arraycopy stub and it could not 245 // be scalar replaced. So we don't check the case. 246 247 // An other case of only one user (Xor) is when the value check for NULL 248 // in G1 post barrier is folded after CCP so the code which used URShift 249 // is removed. 250 251 // Take Region node before eliminating post barrier since it also 252 // eliminates CastP2X node when it has only one user. 253 Node* this_region = p2x->in(0); 254 assert(this_region != NULL, ""); 255 256 // Remove G1 post barrier. 257 258 // Search for CastP2X->Xor->URShift->Cmp path which 259 // checks if the store done to a different from the value's region. 260 // And replace Cmp with #0 (false) to collapse G1 post barrier. 261 Node* xorx = p2x->find_out_with(Op_XorX); 262 assert(xorx != NULL, "missing G1 post barrier"); 263 Node* shift = xorx->unique_out(); 264 Node* cmpx = shift->unique_out(); 265 assert(cmpx->is_Cmp() && cmpx->unique_out()->is_Bool() && 266 cmpx->unique_out()->as_Bool()->_test._test == BoolTest::ne, 267 "missing region check in G1 post barrier"); 268 _igvn.replace_node(cmpx, makecon(TypeInt::CC_EQ)); 269 270 // Remove G1 pre barrier. 271 272 // Search "if (marking != 0)" check and set it to "false". 273 // There is no G1 pre barrier if previous stored value is NULL 274 // (for example, after initialization). 275 if (this_region->is_Region() && this_region->req() == 3) { 276 int ind = 1; 277 if (!this_region->in(ind)->is_IfFalse()) { 278 ind = 2; 279 } 280 if (this_region->in(ind)->is_IfFalse()) { 281 Node* bol = this_region->in(ind)->in(0)->in(1); 282 assert(bol->is_Bool(), ""); 283 cmpx = bol->in(1); 284 if (bol->as_Bool()->_test._test == BoolTest::ne && 285 cmpx->is_Cmp() && cmpx->in(2) == intcon(0) && 286 cmpx->in(1)->is_Load()) { 287 Node* adr = cmpx->in(1)->as_Load()->in(MemNode::Address); 288 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + 289 PtrQueue::byte_offset_of_active()); 290 if (adr->is_AddP() && adr->in(AddPNode::Base) == top() && 291 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal && 292 adr->in(AddPNode::Offset) == MakeConX(marking_offset)) { 293 _igvn.replace_node(cmpx, makecon(TypeInt::CC_EQ)); 294 } 295 } 296 } 297 } 298 // Now CastP2X can be removed since it is used only on dead path 299 // which currently still alive until igvn optimize it. 300 assert(p2x->outcnt() == 0 || p2x->unique_out()->Opcode() == Op_URShiftX, ""); 301 _igvn.replace_node(p2x, top()); 302 } 303 } 304 305 // Search for a memory operation for the specified memory slice. 306 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) { 307 Node *orig_mem = mem; 308 Node *alloc_mem = alloc->in(TypeFunc::Memory); 309 const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr(); 310 while (true) { 311 if (mem == alloc_mem || mem == start_mem ) { 312 return mem; // hit one of our sentinels 313 } else if (mem->is_MergeMem()) { 314 mem = mem->as_MergeMem()->memory_at(alias_idx); 315 } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) { 316 Node *in = mem->in(0); 317 // we can safely skip over safepoints, calls, locks and membars because we 318 // already know that the object is safe to eliminate. 319 if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) { 320 return in; 321 } else if (in->is_Call()) { 322 CallNode *call = in->as_Call(); 323 if (!call->may_modify(tinst, phase)) { 324 mem = call->in(TypeFunc::Memory); 325 } 326 mem = in->in(TypeFunc::Memory); 327 } else if (in->is_MemBar()) { 328 mem = in->in(TypeFunc::Memory); 329 } else { 330 assert(false, "unexpected projection"); 331 } 332 } else if (mem->is_Store()) { 333 const TypePtr* atype = mem->as_Store()->adr_type(); 334 int adr_idx = Compile::current()->get_alias_index(atype); 335 if (adr_idx == alias_idx) { 336 assert(atype->isa_oopptr(), "address type must be oopptr"); 337 int adr_offset = atype->offset(); 338 uint adr_iid = atype->is_oopptr()->instance_id(); 339 // Array elements references have the same alias_idx 340 // but different offset and different instance_id. 341 if (adr_offset == offset && adr_iid == alloc->_idx) 342 return mem; 343 } else { 344 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw"); 345 } 346 mem = mem->in(MemNode::Memory); 347 } else if (mem->is_ClearArray()) { 348 if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) { 349 // Can not bypass initialization of the instance 350 // we are looking. 351 debug_only(intptr_t offset;) 352 assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity"); 353 InitializeNode* init = alloc->as_Allocate()->initialization(); 354 // We are looking for stored value, return Initialize node 355 // or memory edge from Allocate node. 356 if (init != NULL) 357 return init; 358 else 359 return alloc->in(TypeFunc::Memory); // It will produce zero value (see callers). 360 } 361 // Otherwise skip it (the call updated 'mem' value). 362 } else if (mem->Opcode() == Op_SCMemProj) { 363 mem = mem->in(0); 364 Node* adr = NULL; 365 if (mem->is_LoadStore()) { 366 adr = mem->in(MemNode::Address); 367 } else { 368 assert(mem->Opcode() == Op_EncodeISOArray, "sanity"); 369 adr = mem->in(3); // Destination array 370 } 371 const TypePtr* atype = adr->bottom_type()->is_ptr(); 372 int adr_idx = Compile::current()->get_alias_index(atype); 373 if (adr_idx == alias_idx) { 374 assert(false, "Object is not scalar replaceable if a LoadStore node access its field"); 375 return NULL; 376 } 377 mem = mem->in(MemNode::Memory); 378 } else { 379 return mem; 380 } 381 assert(mem != orig_mem, "dead memory loop"); 382 } 383 } 384 385 // 386 // Given a Memory Phi, compute a value Phi containing the values from stores 387 // on the input paths. 388 // Note: this function is recursive, its depth is limied by the "level" argument 389 // Returns the computed Phi, or NULL if it cannot compute it. 390 Node *PhaseMacroExpand::value_from_mem_phi(Node *mem, BasicType ft, const Type *phi_type, const TypeOopPtr *adr_t, Node *alloc, Node_Stack *value_phis, int level) { 391 assert(mem->is_Phi(), "sanity"); 392 int alias_idx = C->get_alias_index(adr_t); 393 int offset = adr_t->offset(); 394 int instance_id = adr_t->instance_id(); 395 396 // Check if an appropriate value phi already exists. 397 Node* region = mem->in(0); 398 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) { 399 Node* phi = region->fast_out(k); 400 if (phi->is_Phi() && phi != mem && 401 phi->as_Phi()->is_same_inst_field(phi_type, instance_id, alias_idx, offset)) { 402 return phi; 403 } 404 } 405 // Check if an appropriate new value phi already exists. 406 Node* new_phi = value_phis->find(mem->_idx); 407 if (new_phi != NULL) 408 return new_phi; 409 410 if (level <= 0) { 411 return NULL; // Give up: phi tree too deep 412 } 413 Node *start_mem = C->start()->proj_out(TypeFunc::Memory); 414 Node *alloc_mem = alloc->in(TypeFunc::Memory); 415 416 uint length = mem->req(); 417 GrowableArray <Node *> values(length, length, NULL, false); 418 419 // create a new Phi for the value 420 PhiNode *phi = new PhiNode(mem->in(0), phi_type, NULL, instance_id, alias_idx, offset); 421 transform_later(phi); 422 value_phis->push(phi, mem->_idx); 423 424 for (uint j = 1; j < length; j++) { 425 Node *in = mem->in(j); 426 if (in == NULL || in->is_top()) { 427 values.at_put(j, in); 428 } else { 429 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn); 430 if (val == start_mem || val == alloc_mem) { 431 // hit a sentinel, return appropriate 0 value 432 values.at_put(j, _igvn.zerocon(ft)); 433 continue; 434 } 435 if (val->is_Initialize()) { 436 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn); 437 } 438 if (val == NULL) { 439 return NULL; // can't find a value on this path 440 } 441 if (val == mem) { 442 values.at_put(j, mem); 443 } else if (val->is_Store()) { 444 values.at_put(j, val->in(MemNode::ValueIn)); 445 } else if(val->is_Proj() && val->in(0) == alloc) { 446 values.at_put(j, _igvn.zerocon(ft)); 447 } else if (val->is_Phi()) { 448 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1); 449 if (val == NULL) { 450 return NULL; 451 } 452 values.at_put(j, val); 453 } else if (val->Opcode() == Op_SCMemProj) { 454 assert(val->in(0)->is_LoadStore() || val->in(0)->Opcode() == Op_EncodeISOArray, "sanity"); 455 assert(false, "Object is not scalar replaceable if a LoadStore node access its field"); 456 return NULL; 457 } else { 458 #ifdef ASSERT 459 val->dump(); 460 assert(false, "unknown node on this path"); 461 #endif 462 return NULL; // unknown node on this path 463 } 464 } 465 } 466 // Set Phi's inputs 467 for (uint j = 1; j < length; j++) { 468 if (values.at(j) == mem) { 469 phi->init_req(j, phi); 470 } else { 471 phi->init_req(j, values.at(j)); 472 } 473 } 474 return phi; 475 } 476 477 // Search the last value stored into the object's field. 478 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, Node *alloc) { 479 assert(adr_t->is_known_instance_field(), "instance required"); 480 int instance_id = adr_t->instance_id(); 481 assert((uint)instance_id == alloc->_idx, "wrong allocation"); 482 483 int alias_idx = C->get_alias_index(adr_t); 484 int offset = adr_t->offset(); 485 Node *start_mem = C->start()->proj_out(TypeFunc::Memory); 486 Node *alloc_ctrl = alloc->in(TypeFunc::Control); 487 Node *alloc_mem = alloc->in(TypeFunc::Memory); 488 Arena *a = Thread::current()->resource_area(); 489 VectorSet visited(a); 490 491 492 bool done = sfpt_mem == alloc_mem; 493 Node *mem = sfpt_mem; 494 while (!done) { 495 if (visited.test_set(mem->_idx)) { 496 return NULL; // found a loop, give up 497 } 498 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn); 499 if (mem == start_mem || mem == alloc_mem) { 500 done = true; // hit a sentinel, return appropriate 0 value 501 } else if (mem->is_Initialize()) { 502 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn); 503 if (mem == NULL) { 504 done = true; // Something go wrong. 505 } else if (mem->is_Store()) { 506 const TypePtr* atype = mem->as_Store()->adr_type(); 507 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice"); 508 done = true; 509 } 510 } else if (mem->is_Store()) { 511 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr(); 512 assert(atype != NULL, "address type must be oopptr"); 513 assert(C->get_alias_index(atype) == alias_idx && 514 atype->is_known_instance_field() && atype->offset() == offset && 515 atype->instance_id() == instance_id, "store is correct memory slice"); 516 done = true; 517 } else if (mem->is_Phi()) { 518 // try to find a phi's unique input 519 Node *unique_input = NULL; 520 Node *top = C->top(); 521 for (uint i = 1; i < mem->req(); i++) { 522 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn); 523 if (n == NULL || n == top || n == mem) { 524 continue; 525 } else if (unique_input == NULL) { 526 unique_input = n; 527 } else if (unique_input != n) { 528 unique_input = top; 529 break; 530 } 531 } 532 if (unique_input != NULL && unique_input != top) { 533 mem = unique_input; 534 } else { 535 done = true; 536 } 537 } else { 538 assert(false, "unexpected node"); 539 } 540 } 541 if (mem != NULL) { 542 if (mem == start_mem || mem == alloc_mem) { 543 // hit a sentinel, return appropriate 0 value 544 return _igvn.zerocon(ft); 545 } else if (mem->is_Store()) { 546 return mem->in(MemNode::ValueIn); 547 } else if (mem->is_Phi()) { 548 // attempt to produce a Phi reflecting the values on the input paths of the Phi 549 Node_Stack value_phis(a, 8); 550 Node * phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit); 551 if (phi != NULL) { 552 return phi; 553 } else { 554 // Kill all new Phis 555 while(value_phis.is_nonempty()) { 556 Node* n = value_phis.node(); 557 _igvn.replace_node(n, C->top()); 558 value_phis.pop(); 559 } 560 } 561 } 562 } 563 // Something go wrong. 564 return NULL; 565 } 566 567 // Check the possibility of scalar replacement. 568 bool PhaseMacroExpand::can_eliminate_allocation(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) { 569 // Scan the uses of the allocation to check for anything that would 570 // prevent us from eliminating it. 571 NOT_PRODUCT( const char* fail_eliminate = NULL; ) 572 DEBUG_ONLY( Node* disq_node = NULL; ) 573 bool can_eliminate = true; 574 575 Node* res = alloc->result_cast(); 576 const TypeOopPtr* res_type = NULL; 577 if (res == NULL) { 578 // All users were eliminated. 579 } else if (!res->is_CheckCastPP()) { 580 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";) 581 can_eliminate = false; 582 } else { 583 res_type = _igvn.type(res)->isa_oopptr(); 584 if (res_type == NULL) { 585 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";) 586 can_eliminate = false; 587 } else if (res_type->isa_aryptr()) { 588 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1); 589 if (length < 0) { 590 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";) 591 can_eliminate = false; 592 } 593 } 594 } 595 596 if (can_eliminate && res != NULL) { 597 for (DUIterator_Fast jmax, j = res->fast_outs(jmax); 598 j < jmax && can_eliminate; j++) { 599 Node* use = res->fast_out(j); 600 601 if (use->is_AddP()) { 602 const TypePtr* addp_type = _igvn.type(use)->is_ptr(); 603 int offset = addp_type->offset(); 604 605 if (offset == Type::OffsetTop || offset == Type::OffsetBot) { 606 NOT_PRODUCT(fail_eliminate = "Undefined field referrence";) 607 can_eliminate = false; 608 break; 609 } 610 for (DUIterator_Fast kmax, k = use->fast_outs(kmax); 611 k < kmax && can_eliminate; k++) { 612 Node* n = use->fast_out(k); 613 if (!n->is_Store() && n->Opcode() != Op_CastP2X) { 614 DEBUG_ONLY(disq_node = n;) 615 if (n->is_Load() || n->is_LoadStore()) { 616 NOT_PRODUCT(fail_eliminate = "Field load";) 617 } else { 618 NOT_PRODUCT(fail_eliminate = "Not store field referrence";) 619 } 620 can_eliminate = false; 621 } 622 } 623 } else if (use->is_SafePoint()) { 624 SafePointNode* sfpt = use->as_SafePoint(); 625 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) { 626 // Object is passed as argument. 627 DEBUG_ONLY(disq_node = use;) 628 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";) 629 can_eliminate = false; 630 } 631 Node* sfptMem = sfpt->memory(); 632 if (sfptMem == NULL || sfptMem->is_top()) { 633 DEBUG_ONLY(disq_node = use;) 634 NOT_PRODUCT(fail_eliminate = "NULL or TOP memory";) 635 can_eliminate = false; 636 } else { 637 safepoints.append_if_missing(sfpt); 638 } 639 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark 640 if (use->is_Phi()) { 641 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) { 642 NOT_PRODUCT(fail_eliminate = "Object is return value";) 643 } else { 644 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";) 645 } 646 DEBUG_ONLY(disq_node = use;) 647 } else { 648 if (use->Opcode() == Op_Return) { 649 NOT_PRODUCT(fail_eliminate = "Object is return value";) 650 }else { 651 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";) 652 } 653 DEBUG_ONLY(disq_node = use;) 654 } 655 can_eliminate = false; 656 } 657 } 658 } 659 660 #ifndef PRODUCT 661 if (PrintEliminateAllocations) { 662 if (can_eliminate) { 663 tty->print("Scalar "); 664 if (res == NULL) 665 alloc->dump(); 666 else 667 res->dump(); 668 } else if (alloc->_is_scalar_replaceable) { 669 tty->print("NotScalar (%s)", fail_eliminate); 670 if (res == NULL) 671 alloc->dump(); 672 else 673 res->dump(); 674 #ifdef ASSERT 675 if (disq_node != NULL) { 676 tty->print(" >>>> "); 677 disq_node->dump(); 678 } 679 #endif /*ASSERT*/ 680 } 681 } 682 #endif 683 return can_eliminate; 684 } 685 686 // Do scalar replacement. 687 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) { 688 GrowableArray <SafePointNode *> safepoints_done; 689 690 ciKlass* klass = NULL; 691 ciInstanceKlass* iklass = NULL; 692 int nfields = 0; 693 int array_base; 694 int element_size; 695 BasicType basic_elem_type; 696 ciType* elem_type; 697 698 Node* res = alloc->result_cast(); 699 assert(res == NULL || res->is_CheckCastPP(), "unexpected AllocateNode result"); 700 const TypeOopPtr* res_type = NULL; 701 if (res != NULL) { // Could be NULL when there are no users 702 res_type = _igvn.type(res)->isa_oopptr(); 703 } 704 705 if (res != NULL) { 706 klass = res_type->klass(); 707 if (res_type->isa_instptr()) { 708 // find the fields of the class which will be needed for safepoint debug information 709 assert(klass->is_instance_klass(), "must be an instance klass."); 710 iklass = klass->as_instance_klass(); 711 nfields = iklass->nof_nonstatic_fields(); 712 } else { 713 // find the array's elements which will be needed for safepoint debug information 714 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1); 715 assert(klass->is_array_klass() && nfields >= 0, "must be an array klass."); 716 elem_type = klass->as_array_klass()->element_type(); 717 basic_elem_type = elem_type->basic_type(); 718 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type); 719 element_size = type2aelembytes(basic_elem_type); 720 } 721 } 722 // 723 // Process the safepoint uses 724 // 725 while (safepoints.length() > 0) { 726 SafePointNode* sfpt = safepoints.pop(); 727 Node* mem = sfpt->memory(); 728 assert(sfpt->jvms() != NULL, "missed JVMS"); 729 // Fields of scalar objs are referenced only at the end 730 // of regular debuginfo at the last (youngest) JVMS. 731 // Record relative start index. 732 uint first_ind = (sfpt->req() - sfpt->jvms()->scloff()); 733 SafePointScalarObjectNode* sobj = new SafePointScalarObjectNode(res_type, 734 #ifdef ASSERT 735 alloc, 736 #endif 737 first_ind, nfields); 738 sobj->init_req(0, C->root()); 739 transform_later(sobj); 740 741 // Scan object's fields adding an input to the safepoint for each field. 742 for (int j = 0; j < nfields; j++) { 743 intptr_t offset; 744 ciField* field = NULL; 745 if (iklass != NULL) { 746 field = iklass->nonstatic_field_at(j); 747 offset = field->offset(); 748 elem_type = field->type(); 749 basic_elem_type = field->layout_type(); 750 } else { 751 offset = array_base + j * (intptr_t)element_size; 752 } 753 754 const Type *field_type; 755 // The next code is taken from Parse::do_get_xxx(). 756 if (basic_elem_type == T_OBJECT || basic_elem_type == T_ARRAY) { 757 if (!elem_type->is_loaded()) { 758 field_type = TypeInstPtr::BOTTOM; 759 } else if (field != NULL && field->is_constant() && field->is_static()) { 760 // This can happen if the constant oop is non-perm. 761 ciObject* con = field->constant_value().as_object(); 762 // Do not "join" in the previous type; it doesn't add value, 763 // and may yield a vacuous result if the field is of interface type. 764 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr(); 765 assert(field_type != NULL, "field singleton type must be consistent"); 766 } else { 767 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass()); 768 } 769 if (UseCompressedOops) { 770 field_type = field_type->make_narrowoop(); 771 basic_elem_type = T_NARROWOOP; 772 } 773 } else { 774 field_type = Type::get_const_basic_type(basic_elem_type); 775 } 776 777 const TypeOopPtr *field_addr_type = res_type->add_offset(offset)->isa_oopptr(); 778 779 Node *field_val = value_from_mem(mem, basic_elem_type, field_type, field_addr_type, alloc); 780 if (field_val == NULL) { 781 // We weren't able to find a value for this field, 782 // give up on eliminating this allocation. 783 784 // Remove any extra entries we added to the safepoint. 785 uint last = sfpt->req() - 1; 786 for (int k = 0; k < j; k++) { 787 sfpt->del_req(last--); 788 } 789 _igvn._worklist.push(sfpt); 790 // rollback processed safepoints 791 while (safepoints_done.length() > 0) { 792 SafePointNode* sfpt_done = safepoints_done.pop(); 793 // remove any extra entries we added to the safepoint 794 last = sfpt_done->req() - 1; 795 for (int k = 0; k < nfields; k++) { 796 sfpt_done->del_req(last--); 797 } 798 JVMState *jvms = sfpt_done->jvms(); 799 jvms->set_endoff(sfpt_done->req()); 800 // Now make a pass over the debug information replacing any references 801 // to SafePointScalarObjectNode with the allocated object. 802 int start = jvms->debug_start(); 803 int end = jvms->debug_end(); 804 for (int i = start; i < end; i++) { 805 if (sfpt_done->in(i)->is_SafePointScalarObject()) { 806 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject(); 807 if (scobj->first_index(jvms) == sfpt_done->req() && 808 scobj->n_fields() == (uint)nfields) { 809 assert(scobj->alloc() == alloc, "sanity"); 810 sfpt_done->set_req(i, res); 811 } 812 } 813 } 814 _igvn._worklist.push(sfpt_done); 815 } 816 #ifndef PRODUCT 817 if (PrintEliminateAllocations) { 818 if (field != NULL) { 819 tty->print("=== At SafePoint node %d can't find value of Field: ", 820 sfpt->_idx); 821 field->print(); 822 int field_idx = C->get_alias_index(field_addr_type); 823 tty->print(" (alias_idx=%d)", field_idx); 824 } else { // Array's element 825 tty->print("=== At SafePoint node %d can't find value of array element [%d]", 826 sfpt->_idx, j); 827 } 828 tty->print(", which prevents elimination of: "); 829 if (res == NULL) 830 alloc->dump(); 831 else 832 res->dump(); 833 } 834 #endif 835 return false; 836 } 837 if (UseCompressedOops && field_type->isa_narrowoop()) { 838 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation 839 // to be able scalar replace the allocation. 840 if (field_val->is_EncodeP()) { 841 field_val = field_val->in(1); 842 } else { 843 field_val = transform_later(new DecodeNNode(field_val, field_val->get_ptr_type())); 844 } 845 } 846 sfpt->add_req(field_val); 847 } 848 JVMState *jvms = sfpt->jvms(); 849 jvms->set_endoff(sfpt->req()); 850 // Now make a pass over the debug information replacing any references 851 // to the allocated object with "sobj" 852 int start = jvms->debug_start(); 853 int end = jvms->debug_end(); 854 sfpt->replace_edges_in_range(res, sobj, start, end); 855 _igvn._worklist.push(sfpt); 856 safepoints_done.append_if_missing(sfpt); // keep it for rollback 857 } 858 return true; 859 } 860 861 // Process users of eliminated allocation. 862 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc) { 863 Node* res = alloc->result_cast(); 864 if (res != NULL) { 865 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) { 866 Node *use = res->last_out(j); 867 uint oc1 = res->outcnt(); 868 869 if (use->is_AddP()) { 870 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) { 871 Node *n = use->last_out(k); 872 uint oc2 = use->outcnt(); 873 if (n->is_Store()) { 874 #ifdef ASSERT 875 // Verify that there is no dependent MemBarVolatile nodes, 876 // they should be removed during IGVN, see MemBarNode::Ideal(). 877 for (DUIterator_Fast pmax, p = n->fast_outs(pmax); 878 p < pmax; p++) { 879 Node* mb = n->fast_out(p); 880 assert(mb->is_Initialize() || !mb->is_MemBar() || 881 mb->req() <= MemBarNode::Precedent || 882 mb->in(MemBarNode::Precedent) != n, 883 "MemBarVolatile should be eliminated for non-escaping object"); 884 } 885 #endif 886 _igvn.replace_node(n, n->in(MemNode::Memory)); 887 } else { 888 eliminate_card_mark(n); 889 } 890 k -= (oc2 - use->outcnt()); 891 } 892 } else { 893 eliminate_card_mark(use); 894 } 895 j -= (oc1 - res->outcnt()); 896 } 897 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted"); 898 _igvn.remove_dead_node(res); 899 } 900 901 // 902 // Process other users of allocation's projections 903 // 904 if (_resproj != NULL && _resproj->outcnt() != 0) { 905 // First disconnect stores captured by Initialize node. 906 // If Initialize node is eliminated first in the following code, 907 // it will kill such stores and DUIterator_Last will assert. 908 for (DUIterator_Fast jmax, j = _resproj->fast_outs(jmax); j < jmax; j++) { 909 Node *use = _resproj->fast_out(j); 910 if (use->is_AddP()) { 911 // raw memory addresses used only by the initialization 912 _igvn.replace_node(use, C->top()); 913 --j; --jmax; 914 } 915 } 916 for (DUIterator_Last jmin, j = _resproj->last_outs(jmin); j >= jmin; ) { 917 Node *use = _resproj->last_out(j); 918 uint oc1 = _resproj->outcnt(); 919 if (use->is_Initialize()) { 920 // Eliminate Initialize node. 921 InitializeNode *init = use->as_Initialize(); 922 assert(init->outcnt() <= 2, "only a control and memory projection expected"); 923 Node *ctrl_proj = init->proj_out(TypeFunc::Control); 924 if (ctrl_proj != NULL) { 925 assert(init->in(TypeFunc::Control) == _fallthroughcatchproj, "allocation control projection"); 926 _igvn.replace_node(ctrl_proj, _fallthroughcatchproj); 927 } 928 Node *mem_proj = init->proj_out(TypeFunc::Memory); 929 if (mem_proj != NULL) { 930 Node *mem = init->in(TypeFunc::Memory); 931 #ifdef ASSERT 932 if (mem->is_MergeMem()) { 933 assert(mem->in(TypeFunc::Memory) == _memproj_fallthrough, "allocation memory projection"); 934 } else { 935 assert(mem == _memproj_fallthrough, "allocation memory projection"); 936 } 937 #endif 938 _igvn.replace_node(mem_proj, mem); 939 } 940 } else { 941 assert(false, "only Initialize or AddP expected"); 942 } 943 j -= (oc1 - _resproj->outcnt()); 944 } 945 } 946 if (_fallthroughcatchproj != NULL) { 947 _igvn.replace_node(_fallthroughcatchproj, alloc->in(TypeFunc::Control)); 948 } 949 if (_memproj_fallthrough != NULL) { 950 _igvn.replace_node(_memproj_fallthrough, alloc->in(TypeFunc::Memory)); 951 } 952 if (_memproj_catchall != NULL) { 953 _igvn.replace_node(_memproj_catchall, C->top()); 954 } 955 if (_ioproj_fallthrough != NULL) { 956 _igvn.replace_node(_ioproj_fallthrough, alloc->in(TypeFunc::I_O)); 957 } 958 if (_ioproj_catchall != NULL) { 959 _igvn.replace_node(_ioproj_catchall, C->top()); 960 } 961 if (_catchallcatchproj != NULL) { 962 _igvn.replace_node(_catchallcatchproj, C->top()); 963 } 964 } 965 966 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) { 967 // Don't do scalar replacement if the frame can be popped by JVMTI: 968 // if reallocation fails during deoptimization we'll pop all 969 // interpreter frames for this compiled frame and that won't play 970 // nice with JVMTI popframe. 971 if (!EliminateAllocations || JvmtiExport::can_pop_frame() || !alloc->_is_non_escaping) { 972 return false; 973 } 974 Node* klass = alloc->in(AllocateNode::KlassNode); 975 const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr(); 976 Node* res = alloc->result_cast(); 977 // Eliminate boxing allocations which are not used 978 // regardless scalar replacable status. 979 bool boxing_alloc = C->eliminate_boxing() && 980 tklass->klass()->is_instance_klass() && 981 tklass->klass()->as_instance_klass()->is_box_klass(); 982 if (!alloc->_is_scalar_replaceable && (!boxing_alloc || (res != NULL))) { 983 return false; 984 } 985 986 extract_call_projections(alloc); 987 988 GrowableArray <SafePointNode *> safepoints; 989 if (!can_eliminate_allocation(alloc, safepoints)) { 990 return false; 991 } 992 993 if (!alloc->_is_scalar_replaceable) { 994 assert(res == NULL, "sanity"); 995 // We can only eliminate allocation if all debug info references 996 // are already replaced with SafePointScalarObject because 997 // we can't search for a fields value without instance_id. 998 if (safepoints.length() > 0) { 999 return false; 1000 } 1001 } 1002 1003 if (!scalar_replacement(alloc, safepoints)) { 1004 return false; 1005 } 1006 1007 CompileLog* log = C->log(); 1008 if (log != NULL) { 1009 log->head("eliminate_allocation type='%d'", 1010 log->identify(tklass->klass())); 1011 JVMState* p = alloc->jvms(); 1012 while (p != NULL) { 1013 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method())); 1014 p = p->caller(); 1015 } 1016 log->tail("eliminate_allocation"); 1017 } 1018 1019 process_users_of_allocation(alloc); 1020 1021 #ifndef PRODUCT 1022 if (PrintEliminateAllocations) { 1023 if (alloc->is_AllocateArray()) 1024 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx); 1025 else 1026 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx); 1027 } 1028 #endif 1029 1030 return true; 1031 } 1032 1033 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) { 1034 // EA should remove all uses of non-escaping boxing node. 1035 if (!C->eliminate_boxing() || boxing->proj_out(TypeFunc::Parms) != NULL) { 1036 return false; 1037 } 1038 1039 assert(boxing->result_cast() == NULL, "unexpected boxing node result"); 1040 1041 extract_call_projections(boxing); 1042 1043 const TypeTuple* r = boxing->tf()->range(); 1044 assert(r->cnt() > TypeFunc::Parms, "sanity"); 1045 const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr(); 1046 assert(t != NULL, "sanity"); 1047 1048 CompileLog* log = C->log(); 1049 if (log != NULL) { 1050 log->head("eliminate_boxing type='%d'", 1051 log->identify(t->klass())); 1052 JVMState* p = boxing->jvms(); 1053 while (p != NULL) { 1054 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method())); 1055 p = p->caller(); 1056 } 1057 log->tail("eliminate_boxing"); 1058 } 1059 1060 process_users_of_allocation(boxing); 1061 1062 #ifndef PRODUCT 1063 if (PrintEliminateAllocations) { 1064 tty->print("++++ Eliminated: %d ", boxing->_idx); 1065 boxing->method()->print_short_name(tty); 1066 tty->cr(); 1067 } 1068 #endif 1069 1070 return true; 1071 } 1072 1073 //---------------------------set_eden_pointers------------------------- 1074 void PhaseMacroExpand::set_eden_pointers(Node* &eden_top_adr, Node* &eden_end_adr) { 1075 if (UseTLAB) { // Private allocation: load from TLS 1076 Node* thread = transform_later(new ThreadLocalNode()); 1077 int tlab_top_offset = in_bytes(JavaThread::tlab_top_offset()); 1078 int tlab_end_offset = in_bytes(JavaThread::tlab_end_offset()); 1079 eden_top_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_top_offset); 1080 eden_end_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_end_offset); 1081 } else { // Shared allocation: load from globals 1082 CollectedHeap* ch = Universe::heap(); 1083 address top_adr = (address)ch->top_addr(); 1084 address end_adr = (address)ch->end_addr(); 1085 eden_top_adr = makecon(TypeRawPtr::make(top_adr)); 1086 eden_end_adr = basic_plus_adr(eden_top_adr, end_adr - top_adr); 1087 } 1088 } 1089 1090 1091 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) { 1092 Node* adr = basic_plus_adr(base, offset); 1093 const TypePtr* adr_type = adr->bottom_type()->is_ptr(); 1094 Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt, MemNode::unordered); 1095 transform_later(value); 1096 return value; 1097 } 1098 1099 1100 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) { 1101 Node* adr = basic_plus_adr(base, offset); 1102 mem = StoreNode::make(_igvn, ctl, mem, adr, NULL, value, bt, MemNode::unordered); 1103 transform_later(mem); 1104 return mem; 1105 } 1106 1107 //============================================================================= 1108 // 1109 // A L L O C A T I O N 1110 // 1111 // Allocation attempts to be fast in the case of frequent small objects. 1112 // It breaks down like this: 1113 // 1114 // 1) Size in doublewords is computed. This is a constant for objects and 1115 // variable for most arrays. Doubleword units are used to avoid size 1116 // overflow of huge doubleword arrays. We need doublewords in the end for 1117 // rounding. 1118 // 1119 // 2) Size is checked for being 'too large'. Too-large allocations will go 1120 // the slow path into the VM. The slow path can throw any required 1121 // exceptions, and does all the special checks for very large arrays. The 1122 // size test can constant-fold away for objects. For objects with 1123 // finalizers it constant-folds the otherway: you always go slow with 1124 // finalizers. 1125 // 1126 // 3) If NOT using TLABs, this is the contended loop-back point. 1127 // Load-Locked the heap top. If using TLABs normal-load the heap top. 1128 // 1129 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route. 1130 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish 1131 // "size*8" we always enter the VM, where "largish" is a constant picked small 1132 // enough that there's always space between the eden max and 4Gig (old space is 1133 // there so it's quite large) and large enough that the cost of entering the VM 1134 // is dwarfed by the cost to initialize the space. 1135 // 1136 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back 1137 // down. If contended, repeat at step 3. If using TLABs normal-store 1138 // adjusted heap top back down; there is no contention. 1139 // 1140 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark 1141 // fields. 1142 // 1143 // 7) Merge with the slow-path; cast the raw memory pointer to the correct 1144 // oop flavor. 1145 // 1146 //============================================================================= 1147 // FastAllocateSizeLimit value is in DOUBLEWORDS. 1148 // Allocations bigger than this always go the slow route. 1149 // This value must be small enough that allocation attempts that need to 1150 // trigger exceptions go the slow route. Also, it must be small enough so 1151 // that heap_top + size_in_bytes does not wrap around the 4Gig limit. 1152 //=============================================================================j// 1153 // %%% Here is an old comment from parseHelper.cpp; is it outdated? 1154 // The allocator will coalesce int->oop copies away. See comment in 1155 // coalesce.cpp about how this works. It depends critically on the exact 1156 // code shape produced here, so if you are changing this code shape 1157 // make sure the GC info for the heap-top is correct in and around the 1158 // slow-path call. 1159 // 1160 1161 void PhaseMacroExpand::expand_allocate_common( 1162 AllocateNode* alloc, // allocation node to be expanded 1163 Node* length, // array length for an array allocation 1164 const TypeFunc* slow_call_type, // Type of slow call 1165 address slow_call_address // Address of slow call 1166 ) 1167 { 1168 1169 Node* ctrl = alloc->in(TypeFunc::Control); 1170 Node* mem = alloc->in(TypeFunc::Memory); 1171 Node* i_o = alloc->in(TypeFunc::I_O); 1172 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize); 1173 Node* klass_node = alloc->in(AllocateNode::KlassNode); 1174 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest); 1175 1176 assert(ctrl != NULL, "must have control"); 1177 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results. 1178 // they will not be used if "always_slow" is set 1179 enum { slow_result_path = 1, fast_result_path = 2 }; 1180 Node *result_region; 1181 Node *result_phi_rawmem; 1182 Node *result_phi_rawoop; 1183 Node *result_phi_i_o; 1184 1185 // The initial slow comparison is a size check, the comparison 1186 // we want to do is a BoolTest::gt 1187 bool always_slow = false; 1188 int tv = _igvn.find_int_con(initial_slow_test, -1); 1189 if (tv >= 0) { 1190 always_slow = (tv == 1); 1191 initial_slow_test = NULL; 1192 } else { 1193 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn); 1194 } 1195 1196 if (C->env()->dtrace_alloc_probes() || 1197 !UseTLAB && (!Universe::heap()->supports_inline_contig_alloc())) { 1198 // Force slow-path allocation 1199 always_slow = true; 1200 initial_slow_test = NULL; 1201 } 1202 1203 1204 enum { too_big_or_final_path = 1, need_gc_path = 2 }; 1205 Node *slow_region = NULL; 1206 Node *toobig_false = ctrl; 1207 1208 assert (initial_slow_test == NULL || !always_slow, "arguments must be consistent"); 1209 // generate the initial test if necessary 1210 if (initial_slow_test != NULL ) { 1211 slow_region = new RegionNode(3); 1212 1213 // Now make the initial failure test. Usually a too-big test but 1214 // might be a TRUE for finalizers or a fancy class check for 1215 // newInstance0. 1216 IfNode *toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN); 1217 transform_later(toobig_iff); 1218 // Plug the failing-too-big test into the slow-path region 1219 Node *toobig_true = new IfTrueNode( toobig_iff ); 1220 transform_later(toobig_true); 1221 slow_region ->init_req( too_big_or_final_path, toobig_true ); 1222 toobig_false = new IfFalseNode( toobig_iff ); 1223 transform_later(toobig_false); 1224 } else { // No initial test, just fall into next case 1225 toobig_false = ctrl; 1226 debug_only(slow_region = NodeSentinel); 1227 } 1228 1229 Node *slow_mem = mem; // save the current memory state for slow path 1230 // generate the fast allocation code unless we know that the initial test will always go slow 1231 if (!always_slow) { 1232 // Fast path modifies only raw memory. 1233 if (mem->is_MergeMem()) { 1234 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw); 1235 } 1236 1237 Node* eden_top_adr; 1238 Node* eden_end_adr; 1239 1240 set_eden_pointers(eden_top_adr, eden_end_adr); 1241 1242 // Load Eden::end. Loop invariant and hoisted. 1243 // 1244 // Note: We set the control input on "eden_end" and "old_eden_top" when using 1245 // a TLAB to work around a bug where these values were being moved across 1246 // a safepoint. These are not oops, so they cannot be include in the oop 1247 // map, but they can be changed by a GC. The proper way to fix this would 1248 // be to set the raw memory state when generating a SafepointNode. However 1249 // this will require extensive changes to the loop optimization in order to 1250 // prevent a degradation of the optimization. 1251 // See comment in memnode.hpp, around line 227 in class LoadPNode. 1252 Node *eden_end = make_load(ctrl, mem, eden_end_adr, 0, TypeRawPtr::BOTTOM, T_ADDRESS); 1253 1254 // allocate the Region and Phi nodes for the result 1255 result_region = new RegionNode(3); 1256 result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM); 1257 result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM); 1258 result_phi_i_o = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch 1259 1260 // We need a Region for the loop-back contended case. 1261 enum { fall_in_path = 1, contended_loopback_path = 2 }; 1262 Node *contended_region; 1263 Node *contended_phi_rawmem; 1264 if (UseTLAB) { 1265 contended_region = toobig_false; 1266 contended_phi_rawmem = mem; 1267 } else { 1268 contended_region = new RegionNode(3); 1269 contended_phi_rawmem = new PhiNode(contended_region, Type::MEMORY, TypeRawPtr::BOTTOM); 1270 // Now handle the passing-too-big test. We fall into the contended 1271 // loop-back merge point. 1272 contended_region ->init_req(fall_in_path, toobig_false); 1273 contended_phi_rawmem->init_req(fall_in_path, mem); 1274 transform_later(contended_region); 1275 transform_later(contended_phi_rawmem); 1276 } 1277 1278 // Load(-locked) the heap top. 1279 // See note above concerning the control input when using a TLAB 1280 Node *old_eden_top = UseTLAB 1281 ? new LoadPNode (ctrl, contended_phi_rawmem, eden_top_adr, TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM, MemNode::unordered) 1282 : new LoadPLockedNode(contended_region, contended_phi_rawmem, eden_top_adr, MemNode::acquire); 1283 1284 transform_later(old_eden_top); 1285 // Add to heap top to get a new heap top 1286 Node *new_eden_top = new AddPNode(top(), old_eden_top, size_in_bytes); 1287 transform_later(new_eden_top); 1288 // Check for needing a GC; compare against heap end 1289 Node *needgc_cmp = new CmpPNode(new_eden_top, eden_end); 1290 transform_later(needgc_cmp); 1291 Node *needgc_bol = new BoolNode(needgc_cmp, BoolTest::ge); 1292 transform_later(needgc_bol); 1293 IfNode *needgc_iff = new IfNode(contended_region, needgc_bol, PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN); 1294 transform_later(needgc_iff); 1295 1296 // Plug the failing-heap-space-need-gc test into the slow-path region 1297 Node *needgc_true = new IfTrueNode(needgc_iff); 1298 transform_later(needgc_true); 1299 if (initial_slow_test) { 1300 slow_region->init_req(need_gc_path, needgc_true); 1301 // This completes all paths into the slow merge point 1302 transform_later(slow_region); 1303 } else { // No initial slow path needed! 1304 // Just fall from the need-GC path straight into the VM call. 1305 slow_region = needgc_true; 1306 } 1307 // No need for a GC. Setup for the Store-Conditional 1308 Node *needgc_false = new IfFalseNode(needgc_iff); 1309 transform_later(needgc_false); 1310 1311 // Grab regular I/O before optional prefetch may change it. 1312 // Slow-path does no I/O so just set it to the original I/O. 1313 result_phi_i_o->init_req(slow_result_path, i_o); 1314 1315 i_o = prefetch_allocation(i_o, needgc_false, contended_phi_rawmem, 1316 old_eden_top, new_eden_top, length); 1317 1318 // Name successful fast-path variables 1319 Node* fast_oop = old_eden_top; 1320 Node* fast_oop_ctrl; 1321 Node* fast_oop_rawmem; 1322 1323 // Store (-conditional) the modified eden top back down. 1324 // StorePConditional produces flags for a test PLUS a modified raw 1325 // memory state. 1326 if (UseTLAB) { 1327 Node* store_eden_top = 1328 new StorePNode(needgc_false, contended_phi_rawmem, eden_top_adr, 1329 TypeRawPtr::BOTTOM, new_eden_top, MemNode::unordered); 1330 transform_later(store_eden_top); 1331 fast_oop_ctrl = needgc_false; // No contention, so this is the fast path 1332 fast_oop_rawmem = store_eden_top; 1333 } else { 1334 Node* store_eden_top = 1335 new StorePConditionalNode(needgc_false, contended_phi_rawmem, eden_top_adr, 1336 new_eden_top, fast_oop/*old_eden_top*/); 1337 transform_later(store_eden_top); 1338 Node *contention_check = new BoolNode(store_eden_top, BoolTest::ne); 1339 transform_later(contention_check); 1340 store_eden_top = new SCMemProjNode(store_eden_top); 1341 transform_later(store_eden_top); 1342 1343 // If not using TLABs, check to see if there was contention. 1344 IfNode *contention_iff = new IfNode (needgc_false, contention_check, PROB_MIN, COUNT_UNKNOWN); 1345 transform_later(contention_iff); 1346 Node *contention_true = new IfTrueNode(contention_iff); 1347 transform_later(contention_true); 1348 // If contention, loopback and try again. 1349 contended_region->init_req(contended_loopback_path, contention_true); 1350 contended_phi_rawmem->init_req(contended_loopback_path, store_eden_top); 1351 1352 // Fast-path succeeded with no contention! 1353 Node *contention_false = new IfFalseNode(contention_iff); 1354 transform_later(contention_false); 1355 fast_oop_ctrl = contention_false; 1356 1357 // Bump total allocated bytes for this thread 1358 Node* thread = new ThreadLocalNode(); 1359 transform_later(thread); 1360 Node* alloc_bytes_adr = basic_plus_adr(top()/*not oop*/, thread, 1361 in_bytes(JavaThread::allocated_bytes_offset())); 1362 Node* alloc_bytes = make_load(fast_oop_ctrl, store_eden_top, alloc_bytes_adr, 1363 0, TypeLong::LONG, T_LONG); 1364 #ifdef _LP64 1365 Node* alloc_size = size_in_bytes; 1366 #else 1367 Node* alloc_size = new ConvI2LNode(size_in_bytes); 1368 transform_later(alloc_size); 1369 #endif 1370 Node* new_alloc_bytes = new AddLNode(alloc_bytes, alloc_size); 1371 transform_later(new_alloc_bytes); 1372 fast_oop_rawmem = make_store(fast_oop_ctrl, store_eden_top, alloc_bytes_adr, 1373 0, new_alloc_bytes, T_LONG); 1374 } 1375 1376 InitializeNode* init = alloc->initialization(); 1377 fast_oop_rawmem = initialize_object(alloc, 1378 fast_oop_ctrl, fast_oop_rawmem, fast_oop, 1379 klass_node, length, size_in_bytes); 1380 1381 // If initialization is performed by an array copy, any required 1382 // MemBarStoreStore was already added. If the object does not 1383 // escape no need for a MemBarStoreStore. Otherwise we need a 1384 // MemBarStoreStore so that stores that initialize this object 1385 // can't be reordered with a subsequent store that makes this 1386 // object accessible by other threads. 1387 if (init == NULL || (!init->is_complete_with_arraycopy() && !init->does_not_escape())) { 1388 if (init == NULL || init->req() < InitializeNode::RawStores) { 1389 // No InitializeNode or no stores captured by zeroing 1390 // elimination. Simply add the MemBarStoreStore after object 1391 // initialization. 1392 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot); 1393 transform_later(mb); 1394 1395 mb->init_req(TypeFunc::Memory, fast_oop_rawmem); 1396 mb->init_req(TypeFunc::Control, fast_oop_ctrl); 1397 fast_oop_ctrl = new ProjNode(mb,TypeFunc::Control); 1398 transform_later(fast_oop_ctrl); 1399 fast_oop_rawmem = new ProjNode(mb,TypeFunc::Memory); 1400 transform_later(fast_oop_rawmem); 1401 } else { 1402 // Add the MemBarStoreStore after the InitializeNode so that 1403 // all stores performing the initialization that were moved 1404 // before the InitializeNode happen before the storestore 1405 // barrier. 1406 1407 Node* init_ctrl = init->proj_out(TypeFunc::Control); 1408 Node* init_mem = init->proj_out(TypeFunc::Memory); 1409 1410 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot); 1411 transform_later(mb); 1412 1413 Node* ctrl = new ProjNode(init,TypeFunc::Control); 1414 transform_later(ctrl); 1415 Node* mem = new ProjNode(init,TypeFunc::Memory); 1416 transform_later(mem); 1417 1418 // The MemBarStoreStore depends on control and memory coming 1419 // from the InitializeNode 1420 mb->init_req(TypeFunc::Memory, mem); 1421 mb->init_req(TypeFunc::Control, ctrl); 1422 1423 ctrl = new ProjNode(mb,TypeFunc::Control); 1424 transform_later(ctrl); 1425 mem = new ProjNode(mb,TypeFunc::Memory); 1426 transform_later(mem); 1427 1428 // All nodes that depended on the InitializeNode for control 1429 // and memory must now depend on the MemBarNode that itself 1430 // depends on the InitializeNode 1431 _igvn.replace_node(init_ctrl, ctrl); 1432 _igvn.replace_node(init_mem, mem); 1433 } 1434 } 1435 1436 if (C->env()->dtrace_extended_probes()) { 1437 // Slow-path call 1438 int size = TypeFunc::Parms + 2; 1439 CallLeafNode *call = new CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(), 1440 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_object_alloc_base), 1441 "dtrace_object_alloc", 1442 TypeRawPtr::BOTTOM); 1443 1444 // Get base of thread-local storage area 1445 Node* thread = new ThreadLocalNode(); 1446 transform_later(thread); 1447 1448 call->init_req(TypeFunc::Parms+0, thread); 1449 call->init_req(TypeFunc::Parms+1, fast_oop); 1450 call->init_req(TypeFunc::Control, fast_oop_ctrl); 1451 call->init_req(TypeFunc::I_O , top()); // does no i/o 1452 call->init_req(TypeFunc::Memory , fast_oop_rawmem); 1453 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr)); 1454 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr)); 1455 transform_later(call); 1456 fast_oop_ctrl = new ProjNode(call,TypeFunc::Control); 1457 transform_later(fast_oop_ctrl); 1458 fast_oop_rawmem = new ProjNode(call,TypeFunc::Memory); 1459 transform_later(fast_oop_rawmem); 1460 } 1461 1462 // Plug in the successful fast-path into the result merge point 1463 result_region ->init_req(fast_result_path, fast_oop_ctrl); 1464 result_phi_rawoop->init_req(fast_result_path, fast_oop); 1465 result_phi_i_o ->init_req(fast_result_path, i_o); 1466 result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem); 1467 } else { 1468 slow_region = ctrl; 1469 result_phi_i_o = i_o; // Rename it to use in the following code. 1470 } 1471 1472 // Generate slow-path call 1473 CallNode *call = new CallStaticJavaNode(slow_call_type, slow_call_address, 1474 OptoRuntime::stub_name(slow_call_address), 1475 alloc->jvms()->bci(), 1476 TypePtr::BOTTOM); 1477 call->init_req( TypeFunc::Control, slow_region ); 1478 call->init_req( TypeFunc::I_O , top() ) ; // does no i/o 1479 call->init_req( TypeFunc::Memory , slow_mem ); // may gc ptrs 1480 call->init_req( TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr) ); 1481 call->init_req( TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr) ); 1482 1483 call->init_req(TypeFunc::Parms+0, klass_node); 1484 if (length != NULL) { 1485 call->init_req(TypeFunc::Parms+1, length); 1486 } 1487 1488 // Copy debug information and adjust JVMState information, then replace 1489 // allocate node with the call 1490 copy_call_debug_info((CallNode *) alloc, call); 1491 if (!always_slow) { 1492 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON. 1493 } else { 1494 // Hook i_o projection to avoid its elimination during allocation 1495 // replacement (when only a slow call is generated). 1496 call->set_req(TypeFunc::I_O, result_phi_i_o); 1497 } 1498 _igvn.replace_node(alloc, call); 1499 transform_later(call); 1500 1501 // Identify the output projections from the allocate node and 1502 // adjust any references to them. 1503 // The control and io projections look like: 1504 // 1505 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl) 1506 // Allocate Catch 1507 // ^---Proj(io) <-------+ ^---CatchProj(io) 1508 // 1509 // We are interested in the CatchProj nodes. 1510 // 1511 extract_call_projections(call); 1512 1513 // An allocate node has separate memory projections for the uses on 1514 // the control and i_o paths. Replace the control memory projection with 1515 // result_phi_rawmem (unless we are only generating a slow call when 1516 // both memory projections are combined) 1517 if (!always_slow && _memproj_fallthrough != NULL) { 1518 for (DUIterator_Fast imax, i = _memproj_fallthrough->fast_outs(imax); i < imax; i++) { 1519 Node *use = _memproj_fallthrough->fast_out(i); 1520 _igvn.rehash_node_delayed(use); 1521 imax -= replace_input(use, _memproj_fallthrough, result_phi_rawmem); 1522 // back up iterator 1523 --i; 1524 } 1525 } 1526 // Now change uses of _memproj_catchall to use _memproj_fallthrough and delete 1527 // _memproj_catchall so we end up with a call that has only 1 memory projection. 1528 if (_memproj_catchall != NULL ) { 1529 if (_memproj_fallthrough == NULL) { 1530 _memproj_fallthrough = new ProjNode(call, TypeFunc::Memory); 1531 transform_later(_memproj_fallthrough); 1532 } 1533 for (DUIterator_Fast imax, i = _memproj_catchall->fast_outs(imax); i < imax; i++) { 1534 Node *use = _memproj_catchall->fast_out(i); 1535 _igvn.rehash_node_delayed(use); 1536 imax -= replace_input(use, _memproj_catchall, _memproj_fallthrough); 1537 // back up iterator 1538 --i; 1539 } 1540 assert(_memproj_catchall->outcnt() == 0, "all uses must be deleted"); 1541 _igvn.remove_dead_node(_memproj_catchall); 1542 } 1543 1544 // An allocate node has separate i_o projections for the uses on the control 1545 // and i_o paths. Always replace the control i_o projection with result i_o 1546 // otherwise incoming i_o become dead when only a slow call is generated 1547 // (it is different from memory projections where both projections are 1548 // combined in such case). 1549 if (_ioproj_fallthrough != NULL) { 1550 for (DUIterator_Fast imax, i = _ioproj_fallthrough->fast_outs(imax); i < imax; i++) { 1551 Node *use = _ioproj_fallthrough->fast_out(i); 1552 _igvn.rehash_node_delayed(use); 1553 imax -= replace_input(use, _ioproj_fallthrough, result_phi_i_o); 1554 // back up iterator 1555 --i; 1556 } 1557 } 1558 // Now change uses of _ioproj_catchall to use _ioproj_fallthrough and delete 1559 // _ioproj_catchall so we end up with a call that has only 1 i_o projection. 1560 if (_ioproj_catchall != NULL ) { 1561 if (_ioproj_fallthrough == NULL) { 1562 _ioproj_fallthrough = new ProjNode(call, TypeFunc::I_O); 1563 transform_later(_ioproj_fallthrough); 1564 } 1565 for (DUIterator_Fast imax, i = _ioproj_catchall->fast_outs(imax); i < imax; i++) { 1566 Node *use = _ioproj_catchall->fast_out(i); 1567 _igvn.rehash_node_delayed(use); 1568 imax -= replace_input(use, _ioproj_catchall, _ioproj_fallthrough); 1569 // back up iterator 1570 --i; 1571 } 1572 assert(_ioproj_catchall->outcnt() == 0, "all uses must be deleted"); 1573 _igvn.remove_dead_node(_ioproj_catchall); 1574 } 1575 1576 // if we generated only a slow call, we are done 1577 if (always_slow) { 1578 // Now we can unhook i_o. 1579 if (result_phi_i_o->outcnt() > 1) { 1580 call->set_req(TypeFunc::I_O, top()); 1581 } else { 1582 assert(result_phi_i_o->unique_ctrl_out() == call, ""); 1583 // Case of new array with negative size known during compilation. 1584 // AllocateArrayNode::Ideal() optimization disconnect unreachable 1585 // following code since call to runtime will throw exception. 1586 // As result there will be no users of i_o after the call. 1587 // Leave i_o attached to this call to avoid problems in preceding graph. 1588 } 1589 return; 1590 } 1591 1592 1593 if (_fallthroughcatchproj != NULL) { 1594 ctrl = _fallthroughcatchproj->clone(); 1595 transform_later(ctrl); 1596 _igvn.replace_node(_fallthroughcatchproj, result_region); 1597 } else { 1598 ctrl = top(); 1599 } 1600 Node *slow_result; 1601 if (_resproj == NULL) { 1602 // no uses of the allocation result 1603 slow_result = top(); 1604 } else { 1605 slow_result = _resproj->clone(); 1606 transform_later(slow_result); 1607 _igvn.replace_node(_resproj, result_phi_rawoop); 1608 } 1609 1610 // Plug slow-path into result merge point 1611 result_region ->init_req( slow_result_path, ctrl ); 1612 result_phi_rawoop->init_req( slow_result_path, slow_result); 1613 result_phi_rawmem->init_req( slow_result_path, _memproj_fallthrough ); 1614 transform_later(result_region); 1615 transform_later(result_phi_rawoop); 1616 transform_later(result_phi_rawmem); 1617 transform_later(result_phi_i_o); 1618 // This completes all paths into the result merge point 1619 } 1620 1621 1622 // Helper for PhaseMacroExpand::expand_allocate_common. 1623 // Initializes the newly-allocated storage. 1624 Node* 1625 PhaseMacroExpand::initialize_object(AllocateNode* alloc, 1626 Node* control, Node* rawmem, Node* object, 1627 Node* klass_node, Node* length, 1628 Node* size_in_bytes) { 1629 InitializeNode* init = alloc->initialization(); 1630 // Store the klass & mark bits 1631 Node* mark_node = NULL; 1632 // For now only enable fast locking for non-array types 1633 if (UseBiasedLocking && (length == NULL)) { 1634 mark_node = make_load(control, rawmem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeRawPtr::BOTTOM, T_ADDRESS); 1635 } else { 1636 mark_node = makecon(TypeRawPtr::make((address)markOopDesc::prototype())); 1637 } 1638 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS); 1639 1640 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA); 1641 int header_size = alloc->minimum_header_size(); // conservatively small 1642 1643 // Array length 1644 if (length != NULL) { // Arrays need length field 1645 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT); 1646 // conservatively small header size: 1647 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1648 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass(); 1649 if (k->is_array_klass()) // we know the exact header size in most cases: 1650 header_size = Klass::layout_helper_header_size(k->layout_helper()); 1651 } 1652 1653 // Clear the object body, if necessary. 1654 if (init == NULL) { 1655 // The init has somehow disappeared; be cautious and clear everything. 1656 // 1657 // This can happen if a node is allocated but an uncommon trap occurs 1658 // immediately. In this case, the Initialize gets associated with the 1659 // trap, and may be placed in a different (outer) loop, if the Allocate 1660 // is in a loop. If (this is rare) the inner loop gets unrolled, then 1661 // there can be two Allocates to one Initialize. The answer in all these 1662 // edge cases is safety first. It is always safe to clear immediately 1663 // within an Allocate, and then (maybe or maybe not) clear some more later. 1664 if (!ZeroTLAB) 1665 rawmem = ClearArrayNode::clear_memory(control, rawmem, object, 1666 header_size, size_in_bytes, 1667 &_igvn); 1668 } else { 1669 if (!init->is_complete()) { 1670 // Try to win by zeroing only what the init does not store. 1671 // We can also try to do some peephole optimizations, 1672 // such as combining some adjacent subword stores. 1673 rawmem = init->complete_stores(control, rawmem, object, 1674 header_size, size_in_bytes, &_igvn); 1675 } 1676 // We have no more use for this link, since the AllocateNode goes away: 1677 init->set_req(InitializeNode::RawAddress, top()); 1678 // (If we keep the link, it just confuses the register allocator, 1679 // who thinks he sees a real use of the address by the membar.) 1680 } 1681 1682 return rawmem; 1683 } 1684 1685 // Generate prefetch instructions for next allocations. 1686 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false, 1687 Node*& contended_phi_rawmem, 1688 Node* old_eden_top, Node* new_eden_top, 1689 Node* length) { 1690 enum { fall_in_path = 1, pf_path = 2 }; 1691 if( UseTLAB && AllocatePrefetchStyle == 2 ) { 1692 // Generate prefetch allocation with watermark check. 1693 // As an allocation hits the watermark, we will prefetch starting 1694 // at a "distance" away from watermark. 1695 1696 Node *pf_region = new RegionNode(3); 1697 Node *pf_phi_rawmem = new PhiNode( pf_region, Type::MEMORY, 1698 TypeRawPtr::BOTTOM ); 1699 // I/O is used for Prefetch 1700 Node *pf_phi_abio = new PhiNode( pf_region, Type::ABIO ); 1701 1702 Node *thread = new ThreadLocalNode(); 1703 transform_later(thread); 1704 1705 Node *eden_pf_adr = new AddPNode( top()/*not oop*/, thread, 1706 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) ); 1707 transform_later(eden_pf_adr); 1708 1709 Node *old_pf_wm = new LoadPNode(needgc_false, 1710 contended_phi_rawmem, eden_pf_adr, 1711 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM, 1712 MemNode::unordered); 1713 transform_later(old_pf_wm); 1714 1715 // check against new_eden_top 1716 Node *need_pf_cmp = new CmpPNode( new_eden_top, old_pf_wm ); 1717 transform_later(need_pf_cmp); 1718 Node *need_pf_bol = new BoolNode( need_pf_cmp, BoolTest::ge ); 1719 transform_later(need_pf_bol); 1720 IfNode *need_pf_iff = new IfNode( needgc_false, need_pf_bol, 1721 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN ); 1722 transform_later(need_pf_iff); 1723 1724 // true node, add prefetchdistance 1725 Node *need_pf_true = new IfTrueNode( need_pf_iff ); 1726 transform_later(need_pf_true); 1727 1728 Node *need_pf_false = new IfFalseNode( need_pf_iff ); 1729 transform_later(need_pf_false); 1730 1731 Node *new_pf_wmt = new AddPNode( top(), old_pf_wm, 1732 _igvn.MakeConX(AllocatePrefetchDistance) ); 1733 transform_later(new_pf_wmt ); 1734 new_pf_wmt->set_req(0, need_pf_true); 1735 1736 Node *store_new_wmt = new StorePNode(need_pf_true, 1737 contended_phi_rawmem, eden_pf_adr, 1738 TypeRawPtr::BOTTOM, new_pf_wmt, 1739 MemNode::unordered); 1740 transform_later(store_new_wmt); 1741 1742 // adding prefetches 1743 pf_phi_abio->init_req( fall_in_path, i_o ); 1744 1745 Node *prefetch_adr; 1746 Node *prefetch; 1747 uint lines = AllocatePrefetchDistance / AllocatePrefetchStepSize; 1748 uint step_size = AllocatePrefetchStepSize; 1749 uint distance = 0; 1750 1751 for ( uint i = 0; i < lines; i++ ) { 1752 prefetch_adr = new AddPNode( old_pf_wm, new_pf_wmt, 1753 _igvn.MakeConX(distance) ); 1754 transform_later(prefetch_adr); 1755 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr ); 1756 transform_later(prefetch); 1757 distance += step_size; 1758 i_o = prefetch; 1759 } 1760 pf_phi_abio->set_req( pf_path, i_o ); 1761 1762 pf_region->init_req( fall_in_path, need_pf_false ); 1763 pf_region->init_req( pf_path, need_pf_true ); 1764 1765 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem ); 1766 pf_phi_rawmem->init_req( pf_path, store_new_wmt ); 1767 1768 transform_later(pf_region); 1769 transform_later(pf_phi_rawmem); 1770 transform_later(pf_phi_abio); 1771 1772 needgc_false = pf_region; 1773 contended_phi_rawmem = pf_phi_rawmem; 1774 i_o = pf_phi_abio; 1775 } else if( UseTLAB && AllocatePrefetchStyle == 3 ) { 1776 // Insert a prefetch for each allocation. 1777 // This code is used for Sparc with BIS. 1778 Node *pf_region = new RegionNode(3); 1779 Node *pf_phi_rawmem = new PhiNode( pf_region, Type::MEMORY, 1780 TypeRawPtr::BOTTOM ); 1781 transform_later(pf_region); 1782 1783 // Generate several prefetch instructions. 1784 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines; 1785 uint step_size = AllocatePrefetchStepSize; 1786 uint distance = AllocatePrefetchDistance; 1787 1788 // Next cache address. 1789 Node *cache_adr = new AddPNode(old_eden_top, old_eden_top, 1790 _igvn.MakeConX(distance)); 1791 transform_later(cache_adr); 1792 cache_adr = new CastP2XNode(needgc_false, cache_adr); 1793 transform_later(cache_adr); 1794 Node* mask = _igvn.MakeConX(~(intptr_t)(step_size-1)); 1795 cache_adr = new AndXNode(cache_adr, mask); 1796 transform_later(cache_adr); 1797 cache_adr = new CastX2PNode(cache_adr); 1798 transform_later(cache_adr); 1799 1800 // Prefetch 1801 Node *prefetch = new PrefetchAllocationNode( contended_phi_rawmem, cache_adr ); 1802 prefetch->set_req(0, needgc_false); 1803 transform_later(prefetch); 1804 contended_phi_rawmem = prefetch; 1805 Node *prefetch_adr; 1806 distance = step_size; 1807 for ( uint i = 1; i < lines; i++ ) { 1808 prefetch_adr = new AddPNode( cache_adr, cache_adr, 1809 _igvn.MakeConX(distance) ); 1810 transform_later(prefetch_adr); 1811 prefetch = new PrefetchAllocationNode( contended_phi_rawmem, prefetch_adr ); 1812 transform_later(prefetch); 1813 distance += step_size; 1814 contended_phi_rawmem = prefetch; 1815 } 1816 } else if( AllocatePrefetchStyle > 0 ) { 1817 // Insert a prefetch for each allocation only on the fast-path 1818 Node *prefetch_adr; 1819 Node *prefetch; 1820 // Generate several prefetch instructions. 1821 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines; 1822 uint step_size = AllocatePrefetchStepSize; 1823 uint distance = AllocatePrefetchDistance; 1824 for ( uint i = 0; i < lines; i++ ) { 1825 prefetch_adr = new AddPNode( old_eden_top, new_eden_top, 1826 _igvn.MakeConX(distance) ); 1827 transform_later(prefetch_adr); 1828 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr ); 1829 // Do not let it float too high, since if eden_top == eden_end, 1830 // both might be null. 1831 if( i == 0 ) { // Set control for first prefetch, next follows it 1832 prefetch->init_req(0, needgc_false); 1833 } 1834 transform_later(prefetch); 1835 distance += step_size; 1836 i_o = prefetch; 1837 } 1838 } 1839 return i_o; 1840 } 1841 1842 1843 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) { 1844 expand_allocate_common(alloc, NULL, 1845 OptoRuntime::new_instance_Type(), 1846 OptoRuntime::new_instance_Java()); 1847 } 1848 1849 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) { 1850 Node* length = alloc->in(AllocateNode::ALength); 1851 InitializeNode* init = alloc->initialization(); 1852 Node* klass_node = alloc->in(AllocateNode::KlassNode); 1853 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass(); 1854 address slow_call_address; // Address of slow call 1855 if (init != NULL && init->is_complete_with_arraycopy() && 1856 k->is_type_array_klass()) { 1857 // Don't zero type array during slow allocation in VM since 1858 // it will be initialized later by arraycopy in compiled code. 1859 slow_call_address = OptoRuntime::new_array_nozero_Java(); 1860 } else { 1861 slow_call_address = OptoRuntime::new_array_Java(); 1862 } 1863 expand_allocate_common(alloc, length, 1864 OptoRuntime::new_array_Type(), 1865 slow_call_address); 1866 } 1867 1868 //-------------------mark_eliminated_box---------------------------------- 1869 // 1870 // During EA obj may point to several objects but after few ideal graph 1871 // transformations (CCP) it may point to only one non escaping object 1872 // (but still using phi), corresponding locks and unlocks will be marked 1873 // for elimination. Later obj could be replaced with a new node (new phi) 1874 // and which does not have escape information. And later after some graph 1875 // reshape other locks and unlocks (which were not marked for elimination 1876 // before) are connected to this new obj (phi) but they still will not be 1877 // marked for elimination since new obj has no escape information. 1878 // Mark all associated (same box and obj) lock and unlock nodes for 1879 // elimination if some of them marked already. 1880 void PhaseMacroExpand::mark_eliminated_box(Node* oldbox, Node* obj) { 1881 if (oldbox->as_BoxLock()->is_eliminated()) 1882 return; // This BoxLock node was processed already. 1883 1884 // New implementation (EliminateNestedLocks) has separate BoxLock 1885 // node for each locked region so mark all associated locks/unlocks as 1886 // eliminated even if different objects are referenced in one locked region 1887 // (for example, OSR compilation of nested loop inside locked scope). 1888 if (EliminateNestedLocks || 1889 oldbox->as_BoxLock()->is_simple_lock_region(NULL, obj)) { 1890 // Box is used only in one lock region. Mark this box as eliminated. 1891 _igvn.hash_delete(oldbox); 1892 oldbox->as_BoxLock()->set_eliminated(); // This changes box's hash value 1893 _igvn.hash_insert(oldbox); 1894 1895 for (uint i = 0; i < oldbox->outcnt(); i++) { 1896 Node* u = oldbox->raw_out(i); 1897 if (u->is_AbstractLock() && !u->as_AbstractLock()->is_non_esc_obj()) { 1898 AbstractLockNode* alock = u->as_AbstractLock(); 1899 // Check lock's box since box could be referenced by Lock's debug info. 1900 if (alock->box_node() == oldbox) { 1901 // Mark eliminated all related locks and unlocks. 1902 alock->set_non_esc_obj(); 1903 } 1904 } 1905 } 1906 return; 1907 } 1908 1909 // Create new "eliminated" BoxLock node and use it in monitor debug info 1910 // instead of oldbox for the same object. 1911 BoxLockNode* newbox = oldbox->clone()->as_BoxLock(); 1912 1913 // Note: BoxLock node is marked eliminated only here and it is used 1914 // to indicate that all associated lock and unlock nodes are marked 1915 // for elimination. 1916 newbox->set_eliminated(); 1917 transform_later(newbox); 1918 1919 // Replace old box node with new box for all users of the same object. 1920 for (uint i = 0; i < oldbox->outcnt();) { 1921 bool next_edge = true; 1922 1923 Node* u = oldbox->raw_out(i); 1924 if (u->is_AbstractLock()) { 1925 AbstractLockNode* alock = u->as_AbstractLock(); 1926 if (alock->box_node() == oldbox && alock->obj_node()->eqv_uncast(obj)) { 1927 // Replace Box and mark eliminated all related locks and unlocks. 1928 alock->set_non_esc_obj(); 1929 _igvn.rehash_node_delayed(alock); 1930 alock->set_box_node(newbox); 1931 next_edge = false; 1932 } 1933 } 1934 if (u->is_FastLock() && u->as_FastLock()->obj_node()->eqv_uncast(obj)) { 1935 FastLockNode* flock = u->as_FastLock(); 1936 assert(flock->box_node() == oldbox, "sanity"); 1937 _igvn.rehash_node_delayed(flock); 1938 flock->set_box_node(newbox); 1939 next_edge = false; 1940 } 1941 1942 // Replace old box in monitor debug info. 1943 if (u->is_SafePoint() && u->as_SafePoint()->jvms()) { 1944 SafePointNode* sfn = u->as_SafePoint(); 1945 JVMState* youngest_jvms = sfn->jvms(); 1946 int max_depth = youngest_jvms->depth(); 1947 for (int depth = 1; depth <= max_depth; depth++) { 1948 JVMState* jvms = youngest_jvms->of_depth(depth); 1949 int num_mon = jvms->nof_monitors(); 1950 // Loop over monitors 1951 for (int idx = 0; idx < num_mon; idx++) { 1952 Node* obj_node = sfn->monitor_obj(jvms, idx); 1953 Node* box_node = sfn->monitor_box(jvms, idx); 1954 if (box_node == oldbox && obj_node->eqv_uncast(obj)) { 1955 int j = jvms->monitor_box_offset(idx); 1956 _igvn.replace_input_of(u, j, newbox); 1957 next_edge = false; 1958 } 1959 } 1960 } 1961 } 1962 if (next_edge) i++; 1963 } 1964 } 1965 1966 //-----------------------mark_eliminated_locking_nodes----------------------- 1967 void PhaseMacroExpand::mark_eliminated_locking_nodes(AbstractLockNode *alock) { 1968 if (EliminateNestedLocks) { 1969 if (alock->is_nested()) { 1970 assert(alock->box_node()->as_BoxLock()->is_eliminated(), "sanity"); 1971 return; 1972 } else if (!alock->is_non_esc_obj()) { // Not eliminated or coarsened 1973 // Only Lock node has JVMState needed here. 1974 if (alock->jvms() != NULL && alock->as_Lock()->is_nested_lock_region()) { 1975 // Mark eliminated related nested locks and unlocks. 1976 Node* obj = alock->obj_node(); 1977 BoxLockNode* box_node = alock->box_node()->as_BoxLock(); 1978 assert(!box_node->is_eliminated(), "should not be marked yet"); 1979 // Note: BoxLock node is marked eliminated only here 1980 // and it is used to indicate that all associated lock 1981 // and unlock nodes are marked for elimination. 1982 box_node->set_eliminated(); // Box's hash is always NO_HASH here 1983 for (uint i = 0; i < box_node->outcnt(); i++) { 1984 Node* u = box_node->raw_out(i); 1985 if (u->is_AbstractLock()) { 1986 alock = u->as_AbstractLock(); 1987 if (alock->box_node() == box_node) { 1988 // Verify that this Box is referenced only by related locks. 1989 assert(alock->obj_node()->eqv_uncast(obj), ""); 1990 // Mark all related locks and unlocks. 1991 alock->set_nested(); 1992 } 1993 } 1994 } 1995 } 1996 return; 1997 } 1998 // Process locks for non escaping object 1999 assert(alock->is_non_esc_obj(), ""); 2000 } // EliminateNestedLocks 2001 2002 if (alock->is_non_esc_obj()) { // Lock is used for non escaping object 2003 // Look for all locks of this object and mark them and 2004 // corresponding BoxLock nodes as eliminated. 2005 Node* obj = alock->obj_node(); 2006 for (uint j = 0; j < obj->outcnt(); j++) { 2007 Node* o = obj->raw_out(j); 2008 if (o->is_AbstractLock() && 2009 o->as_AbstractLock()->obj_node()->eqv_uncast(obj)) { 2010 alock = o->as_AbstractLock(); 2011 Node* box = alock->box_node(); 2012 // Replace old box node with new eliminated box for all users 2013 // of the same object and mark related locks as eliminated. 2014 mark_eliminated_box(box, obj); 2015 } 2016 } 2017 } 2018 } 2019 2020 // we have determined that this lock/unlock can be eliminated, we simply 2021 // eliminate the node without expanding it. 2022 // 2023 // Note: The membar's associated with the lock/unlock are currently not 2024 // eliminated. This should be investigated as a future enhancement. 2025 // 2026 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) { 2027 2028 if (!alock->is_eliminated()) { 2029 return false; 2030 } 2031 #ifdef ASSERT 2032 if (!alock->is_coarsened()) { 2033 // Check that new "eliminated" BoxLock node is created. 2034 BoxLockNode* oldbox = alock->box_node()->as_BoxLock(); 2035 assert(oldbox->is_eliminated(), "should be done already"); 2036 } 2037 #endif 2038 CompileLog* log = C->log(); 2039 if (log != NULL) { 2040 log->head("eliminate_lock lock='%d'", 2041 alock->is_Lock()); 2042 JVMState* p = alock->jvms(); 2043 while (p != NULL) { 2044 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method())); 2045 p = p->caller(); 2046 } 2047 log->tail("eliminate_lock"); 2048 } 2049 2050 #ifndef PRODUCT 2051 if (PrintEliminateLocks) { 2052 if (alock->is_Lock()) { 2053 tty->print_cr("++++ Eliminated: %d Lock", alock->_idx); 2054 } else { 2055 tty->print_cr("++++ Eliminated: %d Unlock", alock->_idx); 2056 } 2057 } 2058 #endif 2059 2060 Node* mem = alock->in(TypeFunc::Memory); 2061 Node* ctrl = alock->in(TypeFunc::Control); 2062 2063 extract_call_projections(alock); 2064 // There are 2 projections from the lock. The lock node will 2065 // be deleted when its last use is subsumed below. 2066 assert(alock->outcnt() == 2 && 2067 _fallthroughproj != NULL && 2068 _memproj_fallthrough != NULL, 2069 "Unexpected projections from Lock/Unlock"); 2070 2071 Node* fallthroughproj = _fallthroughproj; 2072 Node* memproj_fallthrough = _memproj_fallthrough; 2073 2074 // The memory projection from a lock/unlock is RawMem 2075 // The input to a Lock is merged memory, so extract its RawMem input 2076 // (unless the MergeMem has been optimized away.) 2077 if (alock->is_Lock()) { 2078 // Seach for MemBarAcquireLock node and delete it also. 2079 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar(); 2080 assert(membar != NULL && membar->Opcode() == Op_MemBarAcquireLock, ""); 2081 Node* ctrlproj = membar->proj_out(TypeFunc::Control); 2082 Node* memproj = membar->proj_out(TypeFunc::Memory); 2083 _igvn.replace_node(ctrlproj, fallthroughproj); 2084 _igvn.replace_node(memproj, memproj_fallthrough); 2085 2086 // Delete FastLock node also if this Lock node is unique user 2087 // (a loop peeling may clone a Lock node). 2088 Node* flock = alock->as_Lock()->fastlock_node(); 2089 if (flock->outcnt() == 1) { 2090 assert(flock->unique_out() == alock, "sanity"); 2091 _igvn.replace_node(flock, top()); 2092 } 2093 } 2094 2095 // Seach for MemBarReleaseLock node and delete it also. 2096 if (alock->is_Unlock() && ctrl != NULL && ctrl->is_Proj() && 2097 ctrl->in(0)->is_MemBar()) { 2098 MemBarNode* membar = ctrl->in(0)->as_MemBar(); 2099 assert(membar->Opcode() == Op_MemBarReleaseLock && 2100 mem->is_Proj() && membar == mem->in(0), ""); 2101 _igvn.replace_node(fallthroughproj, ctrl); 2102 _igvn.replace_node(memproj_fallthrough, mem); 2103 fallthroughproj = ctrl; 2104 memproj_fallthrough = mem; 2105 ctrl = membar->in(TypeFunc::Control); 2106 mem = membar->in(TypeFunc::Memory); 2107 } 2108 2109 _igvn.replace_node(fallthroughproj, ctrl); 2110 _igvn.replace_node(memproj_fallthrough, mem); 2111 return true; 2112 } 2113 2114 2115 //------------------------------expand_lock_node---------------------- 2116 void PhaseMacroExpand::expand_lock_node(LockNode *lock) { 2117 2118 Node* ctrl = lock->in(TypeFunc::Control); 2119 Node* mem = lock->in(TypeFunc::Memory); 2120 Node* obj = lock->obj_node(); 2121 Node* box = lock->box_node(); 2122 Node* flock = lock->fastlock_node(); 2123 2124 assert(!box->as_BoxLock()->is_eliminated(), "sanity"); 2125 2126 // Make the merge point 2127 Node *region; 2128 Node *mem_phi; 2129 Node *slow_path; 2130 2131 if (UseOptoBiasInlining) { 2132 /* 2133 * See the full description in MacroAssembler::biased_locking_enter(). 2134 * 2135 * if( (mark_word & biased_lock_mask) == biased_lock_pattern ) { 2136 * // The object is biased. 2137 * proto_node = klass->prototype_header; 2138 * o_node = thread | proto_node; 2139 * x_node = o_node ^ mark_word; 2140 * if( (x_node & ~age_mask) == 0 ) { // Biased to the current thread ? 2141 * // Done. 2142 * } else { 2143 * if( (x_node & biased_lock_mask) != 0 ) { 2144 * // The klass's prototype header is no longer biased. 2145 * cas(&mark_word, mark_word, proto_node) 2146 * goto cas_lock; 2147 * } else { 2148 * // The klass's prototype header is still biased. 2149 * if( (x_node & epoch_mask) != 0 ) { // Expired epoch? 2150 * old = mark_word; 2151 * new = o_node; 2152 * } else { 2153 * // Different thread or anonymous biased. 2154 * old = mark_word & (epoch_mask | age_mask | biased_lock_mask); 2155 * new = thread | old; 2156 * } 2157 * // Try to rebias. 2158 * if( cas(&mark_word, old, new) == 0 ) { 2159 * // Done. 2160 * } else { 2161 * goto slow_path; // Failed. 2162 * } 2163 * } 2164 * } 2165 * } else { 2166 * // The object is not biased. 2167 * cas_lock: 2168 * if( FastLock(obj) == 0 ) { 2169 * // Done. 2170 * } else { 2171 * slow_path: 2172 * OptoRuntime::complete_monitor_locking_Java(obj); 2173 * } 2174 * } 2175 */ 2176 2177 region = new RegionNode(5); 2178 // create a Phi for the memory state 2179 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 2180 2181 Node* fast_lock_region = new RegionNode(3); 2182 Node* fast_lock_mem_phi = new PhiNode( fast_lock_region, Type::MEMORY, TypeRawPtr::BOTTOM); 2183 2184 // First, check mark word for the biased lock pattern. 2185 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type()); 2186 2187 // Get fast path - mark word has the biased lock pattern. 2188 ctrl = opt_bits_test(ctrl, fast_lock_region, 1, mark_node, 2189 markOopDesc::biased_lock_mask_in_place, 2190 markOopDesc::biased_lock_pattern, true); 2191 // fast_lock_region->in(1) is set to slow path. 2192 fast_lock_mem_phi->init_req(1, mem); 2193 2194 // Now check that the lock is biased to the current thread and has 2195 // the same epoch and bias as Klass::_prototype_header. 2196 2197 // Special-case a fresh allocation to avoid building nodes: 2198 Node* klass_node = AllocateNode::Ideal_klass(obj, &_igvn); 2199 if (klass_node == NULL) { 2200 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes()); 2201 klass_node = transform_later(LoadKlassNode::make(_igvn, NULL, mem, k_adr, _igvn.type(k_adr)->is_ptr())); 2202 #ifdef _LP64 2203 if (UseCompressedClassPointers && klass_node->is_DecodeNKlass()) { 2204 assert(klass_node->in(1)->Opcode() == Op_LoadNKlass, "sanity"); 2205 klass_node->in(1)->init_req(0, ctrl); 2206 } else 2207 #endif 2208 klass_node->init_req(0, ctrl); 2209 } 2210 Node *proto_node = make_load(ctrl, mem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeX_X, TypeX_X->basic_type()); 2211 2212 Node* thread = transform_later(new ThreadLocalNode()); 2213 Node* cast_thread = transform_later(new CastP2XNode(ctrl, thread)); 2214 Node* o_node = transform_later(new OrXNode(cast_thread, proto_node)); 2215 Node* x_node = transform_later(new XorXNode(o_node, mark_node)); 2216 2217 // Get slow path - mark word does NOT match the value. 2218 Node* not_biased_ctrl = opt_bits_test(ctrl, region, 3, x_node, 2219 (~markOopDesc::age_mask_in_place), 0); 2220 // region->in(3) is set to fast path - the object is biased to the current thread. 2221 mem_phi->init_req(3, mem); 2222 2223 2224 // Mark word does NOT match the value (thread | Klass::_prototype_header). 2225 2226 2227 // First, check biased pattern. 2228 // Get fast path - _prototype_header has the same biased lock pattern. 2229 ctrl = opt_bits_test(not_biased_ctrl, fast_lock_region, 2, x_node, 2230 markOopDesc::biased_lock_mask_in_place, 0, true); 2231 2232 not_biased_ctrl = fast_lock_region->in(2); // Slow path 2233 // fast_lock_region->in(2) - the prototype header is no longer biased 2234 // and we have to revoke the bias on this object. 2235 // We are going to try to reset the mark of this object to the prototype 2236 // value and fall through to the CAS-based locking scheme. 2237 Node* adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); 2238 Node* cas = new StoreXConditionalNode(not_biased_ctrl, mem, adr, 2239 proto_node, mark_node); 2240 transform_later(cas); 2241 Node* proj = transform_later(new SCMemProjNode(cas)); 2242 fast_lock_mem_phi->init_req(2, proj); 2243 2244 2245 // Second, check epoch bits. 2246 Node* rebiased_region = new RegionNode(3); 2247 Node* old_phi = new PhiNode( rebiased_region, TypeX_X); 2248 Node* new_phi = new PhiNode( rebiased_region, TypeX_X); 2249 2250 // Get slow path - mark word does NOT match epoch bits. 2251 Node* epoch_ctrl = opt_bits_test(ctrl, rebiased_region, 1, x_node, 2252 markOopDesc::epoch_mask_in_place, 0); 2253 // The epoch of the current bias is not valid, attempt to rebias the object 2254 // toward the current thread. 2255 rebiased_region->init_req(2, epoch_ctrl); 2256 old_phi->init_req(2, mark_node); 2257 new_phi->init_req(2, o_node); 2258 2259 // rebiased_region->in(1) is set to fast path. 2260 // The epoch of the current bias is still valid but we know 2261 // nothing about the owner; it might be set or it might be clear. 2262 Node* cmask = MakeConX(markOopDesc::biased_lock_mask_in_place | 2263 markOopDesc::age_mask_in_place | 2264 markOopDesc::epoch_mask_in_place); 2265 Node* old = transform_later(new AndXNode(mark_node, cmask)); 2266 cast_thread = transform_later(new CastP2XNode(ctrl, thread)); 2267 Node* new_mark = transform_later(new OrXNode(cast_thread, old)); 2268 old_phi->init_req(1, old); 2269 new_phi->init_req(1, new_mark); 2270 2271 transform_later(rebiased_region); 2272 transform_later(old_phi); 2273 transform_later(new_phi); 2274 2275 // Try to acquire the bias of the object using an atomic operation. 2276 // If this fails we will go in to the runtime to revoke the object's bias. 2277 cas = new StoreXConditionalNode(rebiased_region, mem, adr, new_phi, old_phi); 2278 transform_later(cas); 2279 proj = transform_later(new SCMemProjNode(cas)); 2280 2281 // Get slow path - Failed to CAS. 2282 not_biased_ctrl = opt_bits_test(rebiased_region, region, 4, cas, 0, 0); 2283 mem_phi->init_req(4, proj); 2284 // region->in(4) is set to fast path - the object is rebiased to the current thread. 2285 2286 // Failed to CAS. 2287 slow_path = new RegionNode(3); 2288 Node *slow_mem = new PhiNode( slow_path, Type::MEMORY, TypeRawPtr::BOTTOM); 2289 2290 slow_path->init_req(1, not_biased_ctrl); // Capture slow-control 2291 slow_mem->init_req(1, proj); 2292 2293 // Call CAS-based locking scheme (FastLock node). 2294 2295 transform_later(fast_lock_region); 2296 transform_later(fast_lock_mem_phi); 2297 2298 // Get slow path - FastLock failed to lock the object. 2299 ctrl = opt_bits_test(fast_lock_region, region, 2, flock, 0, 0); 2300 mem_phi->init_req(2, fast_lock_mem_phi); 2301 // region->in(2) is set to fast path - the object is locked to the current thread. 2302 2303 slow_path->init_req(2, ctrl); // Capture slow-control 2304 slow_mem->init_req(2, fast_lock_mem_phi); 2305 2306 transform_later(slow_path); 2307 transform_later(slow_mem); 2308 // Reset lock's memory edge. 2309 lock->set_req(TypeFunc::Memory, slow_mem); 2310 2311 } else { 2312 region = new RegionNode(3); 2313 // create a Phi for the memory state 2314 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 2315 2316 // Optimize test; set region slot 2 2317 slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0); 2318 mem_phi->init_req(2, mem); 2319 } 2320 2321 // Make slow path call 2322 CallNode *call = make_slow_call( (CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(), OptoRuntime::complete_monitor_locking_Java(), NULL, slow_path, obj, box ); 2323 2324 extract_call_projections(call); 2325 2326 // Slow path can only throw asynchronous exceptions, which are always 2327 // de-opted. So the compiler thinks the slow-call can never throw an 2328 // exception. If it DOES throw an exception we would need the debug 2329 // info removed first (since if it throws there is no monitor). 2330 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL && 2331 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock"); 2332 2333 // Capture slow path 2334 // disconnect fall-through projection from call and create a new one 2335 // hook up users of fall-through projection to region 2336 Node *slow_ctrl = _fallthroughproj->clone(); 2337 transform_later(slow_ctrl); 2338 _igvn.hash_delete(_fallthroughproj); 2339 _fallthroughproj->disconnect_inputs(NULL, C); 2340 region->init_req(1, slow_ctrl); 2341 // region inputs are now complete 2342 transform_later(region); 2343 _igvn.replace_node(_fallthroughproj, region); 2344 2345 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory)); 2346 mem_phi->init_req(1, memproj ); 2347 transform_later(mem_phi); 2348 _igvn.replace_node(_memproj_fallthrough, mem_phi); 2349 } 2350 2351 //------------------------------expand_unlock_node---------------------- 2352 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) { 2353 2354 Node* ctrl = unlock->in(TypeFunc::Control); 2355 Node* mem = unlock->in(TypeFunc::Memory); 2356 Node* obj = unlock->obj_node(); 2357 Node* box = unlock->box_node(); 2358 2359 assert(!box->as_BoxLock()->is_eliminated(), "sanity"); 2360 2361 // No need for a null check on unlock 2362 2363 // Make the merge point 2364 Node *region; 2365 Node *mem_phi; 2366 2367 if (UseOptoBiasInlining) { 2368 // Check for biased locking unlock case, which is a no-op. 2369 // See the full description in MacroAssembler::biased_locking_exit(). 2370 region = new RegionNode(4); 2371 // create a Phi for the memory state 2372 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 2373 mem_phi->init_req(3, mem); 2374 2375 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type()); 2376 ctrl = opt_bits_test(ctrl, region, 3, mark_node, 2377 markOopDesc::biased_lock_mask_in_place, 2378 markOopDesc::biased_lock_pattern); 2379 } else { 2380 region = new RegionNode(3); 2381 // create a Phi for the memory state 2382 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 2383 } 2384 2385 FastUnlockNode *funlock = new FastUnlockNode( ctrl, obj, box ); 2386 funlock = transform_later( funlock )->as_FastUnlock(); 2387 // Optimize test; set region slot 2 2388 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0); 2389 2390 CallNode *call = make_slow_call( (CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C), "complete_monitor_unlocking_C", slow_path, obj, box ); 2391 2392 extract_call_projections(call); 2393 2394 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL && 2395 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock"); 2396 2397 // No exceptions for unlocking 2398 // Capture slow path 2399 // disconnect fall-through projection from call and create a new one 2400 // hook up users of fall-through projection to region 2401 Node *slow_ctrl = _fallthroughproj->clone(); 2402 transform_later(slow_ctrl); 2403 _igvn.hash_delete(_fallthroughproj); 2404 _fallthroughproj->disconnect_inputs(NULL, C); 2405 region->init_req(1, slow_ctrl); 2406 // region inputs are now complete 2407 transform_later(region); 2408 _igvn.replace_node(_fallthroughproj, region); 2409 2410 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory) ); 2411 mem_phi->init_req(1, memproj ); 2412 mem_phi->init_req(2, mem); 2413 transform_later(mem_phi); 2414 _igvn.replace_node(_memproj_fallthrough, mem_phi); 2415 } 2416 2417 //---------------------------eliminate_macro_nodes---------------------- 2418 // Eliminate scalar replaced allocations and associated locks. 2419 void PhaseMacroExpand::eliminate_macro_nodes() { 2420 if (C->macro_count() == 0) 2421 return; 2422 2423 // First, attempt to eliminate locks 2424 int cnt = C->macro_count(); 2425 for (int i=0; i < cnt; i++) { 2426 Node *n = C->macro_node(i); 2427 if (n->is_AbstractLock()) { // Lock and Unlock nodes 2428 // Before elimination mark all associated (same box and obj) 2429 // lock and unlock nodes. 2430 mark_eliminated_locking_nodes(n->as_AbstractLock()); 2431 } 2432 } 2433 bool progress = true; 2434 while (progress) { 2435 progress = false; 2436 for (int i = C->macro_count(); i > 0; i--) { 2437 Node * n = C->macro_node(i-1); 2438 bool success = false; 2439 debug_only(int old_macro_count = C->macro_count();); 2440 if (n->is_AbstractLock()) { 2441 success = eliminate_locking_node(n->as_AbstractLock()); 2442 } 2443 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count"); 2444 progress = progress || success; 2445 } 2446 } 2447 // Next, attempt to eliminate allocations 2448 _has_locks = false; 2449 progress = true; 2450 while (progress) { 2451 progress = false; 2452 for (int i = C->macro_count(); i > 0; i--) { 2453 Node * n = C->macro_node(i-1); 2454 bool success = false; 2455 debug_only(int old_macro_count = C->macro_count();); 2456 switch (n->class_id()) { 2457 case Node::Class_Allocate: 2458 case Node::Class_AllocateArray: 2459 success = eliminate_allocate_node(n->as_Allocate()); 2460 break; 2461 case Node::Class_CallStaticJava: 2462 success = eliminate_boxing_node(n->as_CallStaticJava()); 2463 break; 2464 case Node::Class_Lock: 2465 case Node::Class_Unlock: 2466 assert(!n->as_AbstractLock()->is_eliminated(), "sanity"); 2467 _has_locks = true; 2468 break; 2469 case Node::Class_ArrayCopy: 2470 break; 2471 default: 2472 assert(n->Opcode() == Op_LoopLimit || 2473 n->Opcode() == Op_Opaque1 || 2474 n->Opcode() == Op_Opaque2 || 2475 n->Opcode() == Op_Opaque3, "unknown node type in macro list"); 2476 } 2477 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count"); 2478 progress = progress || success; 2479 } 2480 } 2481 } 2482 2483 //------------------------------expand_macro_nodes---------------------- 2484 // Returns true if a failure occurred. 2485 bool PhaseMacroExpand::expand_macro_nodes() { 2486 // Last attempt to eliminate macro nodes. 2487 eliminate_macro_nodes(); 2488 2489 // Make sure expansion will not cause node limit to be exceeded. 2490 // Worst case is a macro node gets expanded into about 50 nodes. 2491 // Allow 50% more for optimization. 2492 if (C->check_node_count(C->macro_count() * 75, "out of nodes before macro expansion" ) ) 2493 return true; 2494 2495 // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations. 2496 bool progress = true; 2497 while (progress) { 2498 progress = false; 2499 for (int i = C->macro_count(); i > 0; i--) { 2500 Node * n = C->macro_node(i-1); 2501 bool success = false; 2502 debug_only(int old_macro_count = C->macro_count();); 2503 if (n->Opcode() == Op_LoopLimit) { 2504 // Remove it from macro list and put on IGVN worklist to optimize. 2505 C->remove_macro_node(n); 2506 _igvn._worklist.push(n); 2507 success = true; 2508 } else if (n->Opcode() == Op_CallStaticJava) { 2509 // Remove it from macro list and put on IGVN worklist to optimize. 2510 C->remove_macro_node(n); 2511 _igvn._worklist.push(n); 2512 success = true; 2513 } else if (n->Opcode() == Op_Opaque1 || n->Opcode() == Op_Opaque2) { 2514 _igvn.replace_node(n, n->in(1)); 2515 success = true; 2516 #if INCLUDE_RTM_OPT 2517 } else if ((n->Opcode() == Op_Opaque3) && ((Opaque3Node*)n)->rtm_opt()) { 2518 assert(C->profile_rtm(), "should be used only in rtm deoptimization code"); 2519 assert((n->outcnt() == 1) && n->unique_out()->is_Cmp(), ""); 2520 Node* cmp = n->unique_out(); 2521 #ifdef ASSERT 2522 // Validate graph. 2523 assert((cmp->outcnt() == 1) && cmp->unique_out()->is_Bool(), ""); 2524 BoolNode* bol = cmp->unique_out()->as_Bool(); 2525 assert((bol->outcnt() == 1) && bol->unique_out()->is_If() && 2526 (bol->_test._test == BoolTest::ne), ""); 2527 IfNode* ifn = bol->unique_out()->as_If(); 2528 assert((ifn->outcnt() == 2) && 2529 ifn->proj_out(1)->is_uncommon_trap_proj(Deoptimization::Reason_rtm_state_change), ""); 2530 #endif 2531 Node* repl = n->in(1); 2532 if (!_has_locks) { 2533 // Remove RTM state check if there are no locks in the code. 2534 // Replace input to compare the same value. 2535 repl = (cmp->in(1) == n) ? cmp->in(2) : cmp->in(1); 2536 } 2537 _igvn.replace_node(n, repl); 2538 success = true; 2539 #endif 2540 } 2541 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count"); 2542 progress = progress || success; 2543 } 2544 } 2545 2546 // expand arraycopy "macro" nodes first 2547 // For ReduceBulkZeroing, we must first process all arraycopy nodes 2548 // before the allocate nodes are expanded. 2549 int macro_idx = C->macro_count() - 1; 2550 while (macro_idx >= 0) { 2551 Node * n = C->macro_node(macro_idx); 2552 assert(n->is_macro(), "only macro nodes expected here"); 2553 if (_igvn.type(n) == Type::TOP || n->in(0)->is_top() ) { 2554 // node is unreachable, so don't try to expand it 2555 C->remove_macro_node(n); 2556 } else if (n->is_ArrayCopy()){ 2557 int macro_count = C->macro_count(); 2558 expand_arraycopy_node(n->as_ArrayCopy()); 2559 assert(C->macro_count() < macro_count, "must have deleted a node from macro list"); 2560 } 2561 if (C->failing()) return true; 2562 macro_idx --; 2563 } 2564 2565 // expand "macro" nodes 2566 // nodes are removed from the macro list as they are processed 2567 while (C->macro_count() > 0) { 2568 int macro_count = C->macro_count(); 2569 Node * n = C->macro_node(macro_count-1); 2570 assert(n->is_macro(), "only macro nodes expected here"); 2571 if (_igvn.type(n) == Type::TOP || n->in(0)->is_top() ) { 2572 // node is unreachable, so don't try to expand it 2573 C->remove_macro_node(n); 2574 continue; 2575 } 2576 switch (n->class_id()) { 2577 case Node::Class_Allocate: 2578 expand_allocate(n->as_Allocate()); 2579 break; 2580 case Node::Class_AllocateArray: 2581 expand_allocate_array(n->as_AllocateArray()); 2582 break; 2583 case Node::Class_Lock: 2584 expand_lock_node(n->as_Lock()); 2585 break; 2586 case Node::Class_Unlock: 2587 expand_unlock_node(n->as_Unlock()); 2588 break; 2589 default: 2590 assert(false, "unknown node type in macro list"); 2591 } 2592 assert(C->macro_count() < macro_count, "must have deleted a node from macro list"); 2593 if (C->failing()) return true; 2594 } 2595 2596 _igvn.set_delay_transform(false); 2597 _igvn.optimize(); 2598 if (C->failing()) return true; 2599 return false; 2600 }